sign in sign up
<<
Home , Diffeomorphism, Symplectomorphism

The of the of symplectic diffeomorphisms

Leonid Polterovich

October 15, 2007 Preface

The group of Hamiltonian diffeomorphisms Ham(M, Ω) of a symplectic (M, Ω) plays a fundamental role both in geometry and clas- sical mechanics. For a geometer, at least under some assumptions on the manifold M, this is just the connected component of the identity in the group of all symplectic diffeomorphisms. From the viewpoint of mechanics, Ham(M, Ω) is the group of all admissible motions. What is the minimal amount of energy required in to generate a given Hamiltonian diffeomorphism f? An attempt to formalize and answer this natural question has led H. Hofer [H1] (1990) to a remarkable dis- covery. It turns out that the solution of this variational problem can be interpreted as a geometric quantity, namely as the distance between f and the identity transformation. Moreover this distance is associated to a canonical biinvariant metric on Ham(M, Ω). Since Hofer’s work this new geometry has been intensively studied in the framework of modern symplectic topology. In the present book I will describe some of these developments. Hofer’s geometry enables us to study various notions and prob- lems which come from the familiar finite dimensional geometry in the context of the group of Hamiltonian diffeomorphisms. They turn out to be very different from the usual circle of problems considered in sym- plectic topology and thus extend significantly our vision of the sym- plectic world. Is the diameter of Ham(M, Ω) finite or infinite? What are the minimal geodesics? How can one find the length spectrum? In general, these questions are still open. However some partial answers do exist and will be discussed below. There is one more, to my taste even more important reason why it is useful to have a canonical geometry on the group of Hamiltonian

i ii

diffeomorphisms. Consider a time dependent vector field ξt, t ∈ R on a manifold M. The ordinary differential equation

x˙ = ξ(x, t) on M defines a flow ft : M → M which takes any initial condition x(0) to the solution x(t) at time t. The trajectories of the flow form a complicated system of on the manifold. Usually, in order to understand the dynamics, one should travel along the manifold and thoroughly study behaviour of the trajectories in different regions. Let us change the point of view and note that our flow can be interpreted as a simple geometric object - a single t → ft on the group of all diffeomorphisms of the manifold. One may hope that geometric properties of this curve reflect the dynamics, and thus complicated dynamical phenomena can be studied by purely geometric tools. Of course, the price we pay is that the ambient space - the group of diffeomorphisms - is infinite-dimensional. Moreover, we face a serious difficulty: in general there are no canonical tools to perform geometric measurements on this group. Remarkably, Hofer’s metric provides such a tool for systems of classical mechanics. In Chapters 8 and 11 we will see some situations where this ideology turns out to be successful. As it often happens with very young and fast developing areas of , the proofs of various elegant statements in Hofer’s geometry are technical and complicated. Thus I selected the simplest non-trivial versions of the main phenomena (according to my taste, of course) without making an attempt to present them in the most general form. For the same reason many technicalities are omitted. Though formally speaking this book does not require a special back- ground in symplectic topology (at least necessary definitions and for- mulations are given), the reader is cordially invited to consult two remarkable introductory texts [HZ] and [MS]. Both of them contain chapters on the geometry of the group of Hamiltonian diffeomor- phisms. I have tried to minimize the overlaps. The book contains a number of exercises which presumably will help the reader to get into the subject. This book arose from two sources. The first one is lectures given to graduate students, namely mini-courses at Universities of Freiburg iii and Warwick, and a Nachdiplom-course at ETH, Z¨urich. The second source is the expository article [P8] which contains a brief outline of the material presented below. Let me introduce briefly the main characters of the book. A dif- feomorphism f of a (M, Ω) is called Hamiltonian if it can be included into a compactly supported Hamiltonian flow {ft} with f0 = 1l and f1 = f. Such a flow is defined by a Hamiltonian F : M × [0; 1] → R. In the language of classical mechanics F is the energy of a mechanical described by {ft}. We interpret the total energy of the flow as a length of the corresponding path of diffeomorphisms:

Z 1 length{ft} = max F (x, t) − min F (x, t)dt. 0 x∈M x∈M Define a function

ρ : Ham(M, Ω) × Ham(M, Ω) → R by ρ(φ, ψ) = inf length{ft}, where the infimum is taken over all Hamiltonian flows {ft} which generate the Hamiltonian diffeomorphism f = φψ−1. It is easy to see that ρ is non-negative, symmetric, vanishes on the diagonal and satisfies the triangle inequality. Moreover, ρ is biinvariant with respect to the group structure of Ham(M, Ω). In other words, ρ is a biinvariant pseudo-distance. It is a deep fact that ρ is a genuine distance function, that is ρ(φ, ψ) is strictly positive for φ 6= ψ. The metric ρ is called Hofer’s metric. The group Ham(M, Ω) and Hofer’s pseudo-distance ρ are intro- duced in Chapters 1 and 2 respectively. In Chapter 3 we prove that ρ is a genuine metric in the case when M is the standard symplectic linear space R2n. Our approach is based on Gromov’s theory of holomor- phic discs with Lagrangian boundary conditions which is presented in Chapter 4. Afterwards we turn to the study of basic geometric invariants of Ham(M, Ω). There is a (still open!) that the diameter of iv

Ham(M, Ω) is infinite. In Chapters 5-7 we prove this conjecture for closed surfaces. In Chapter 8 we discuss the concept of the growth of a one- parameter {ft} of Ham(M, Ω) which reflects the asymptotic behaviour of the function ρ(1l, ft) when t → ∞. We present a link be- tween the growth and the dynamics of {ft} in the context of invariant tori of classical mechanics. In many interesting situations the space Ham(M, Ω) has compli- cated topology, and in particular a non-trivial . For an element γ ∈ π1(Ham(M, Ω)) ν(γ) = inf length{ft}, where the infimum is taken over all loops {ft} of Hamiltonian diffeomorphisms (that is periodic Hamiltonian flows) which represent γ. The set ¯ ¯ {ν(γ) ¯ γ ∈ π1(Ham(M, Ω))} is called the length spectrum of Ham(M, Ω). In Chapter 9 we present an approach to the length spectrum which is based on the theory of symplectic fibrations. An important ingredient of this approach is Gromov’s theory of pseudo-holomorphic curves which is discussed in Chapter 10. In Chapter 11 we give an application of our results on the length spectrum to the classical ergodic theory. In Chapters 12 and 13 we develop two different approaches to the theory of geodesics on Ham(M, Ω). One of them is elementary, while the other requires a powerful machinery of . Chapter 13 suggests to the reader a brief visit to Floer homology. Finally, in Chapter 14 we deal with non-Hamiltonian symplectic diffeomorphisms, which appear naturally in Hofer’s geometry as isome- tries of Ham(M, Ω). In addition, we formulate and discuss the famous flux conjecture which states that the group Ham(M, Ω) is closed in the group of all symplectic diffeomorphisms endowed with the C∞- topology. v

Acknowledgments. I cordially thank Meike Akveld for her in- dispensable help in typing the preliminary version of the manuscript, preparing the pictures and enormous editorial work. I am very grateful to Paul Biran and Karl Friedrich Siburg for their detailed comments on the manuscript and for improving the presentation. I am indebted to Rami Aizenbud, Dima Gourevitch, Misha Entov, Osya Polterovich and Zeev Rudnick for pointing out a number of inaccuracies in the pre- liminary version of the book. The book was written during my stay at ETH, Zurich in the academic year 1997-1998, and during my visits to IHES, Bures-sur-Yvette in 1998 and 1999. I thank both institutions for the excellent research atmosphere. vi Contents

1 Introducing the group 1 1.1 The origins of Hamiltonian diffeomorphisms ...... 1 1.2 Flows and paths of diffeomorphisms ...... 4 1.3 Classical mechanics ...... 5 1.4 The group of Hamiltonian diffeomorphisms ...... 7 1.5 Algebraic properties of Ham(M, Ω) ...... 13

2 Introducing the geometry 15 2.1 A variational problem ...... 15 2.2 Biinvariant on Ham(M, Ω) ...... 16 2.3 The choice of the norm: Lp vs. L∞ ...... 18 2.4 The concept of displacement energy ...... 19

3 Lagrangian 25 3.1 Definitions and examples ...... 25 3.2 The Liouville class ...... 27 3.3 Estimating the displacement energy ...... 31

4 The ∂¯-equation 35 4.1 Introducing the ∂¯-operator ...... 35 4.2 The boundary value problem ...... 37 4.3 An application to the Liouville class ...... 39 4.4 An example ...... 41

5 Linearization 45 5.1 The space of periodic Hamiltonians ...... 45 5.2 Regularization ...... 48

vii viii CONTENTS

5.3 Paths in a given homotopy class ...... 50

6 Lagrangian intersections 53 6.1 Exact Lagrangian isotopies ...... 53 6.2 Lagrangian intersections ...... 57 6.3 An application to Hamiltonian loops ...... 58

7 Diameter 61 7.1 The starting estimate ...... 61 7.2 The fundamental group ...... 62 7.3 The length spectrum ...... 65 7.4 Refining the estimate ...... 66

8 Growth and dynamics 69 8.1 Invariant tori of classical mechanics ...... 69 8.2 Growth of one-parameter ...... 72 8.3 Curve shortening in Hofer’s geometry ...... 76 8.4 What happens when the asymptotic growth vanishes? . 78

9 Length spectrum 79 9.1 The positive and negative parts of Hofer’s norm . . . . 79 9.2 Symplectic fibrations over S2 ...... 80 9.3 Symplectic connections ...... 83 9.4 An application to length spectrum ...... 90

10 Deformations of symplectic forms 93 10.1 The deformation problem ...... 93 10.2 The ∂¯-equation revisited ...... 94 10.3 An application to coupling ...... 97 10.4 Pseudo-holomorphic curves ...... 98 10.5 Persistence of exceptional ...... 100

11 Ergodic Theory 103 11.1 Hamiltonian loops as dynamical objects ...... 103 11.2 The asymptotic length spectrum ...... 105 11.3 Geometry via algebra ...... 108 CONTENTS ix

12 Geodesics 111 12.1 What are geodesics? ...... 111 12.2 Description of geodesics ...... 114 12.3 Stability and conjugate points ...... 116 12.4 The second variation formula ...... 117 12.5 Analysis of the second variation formula ...... 123 12.6 Length minimizing geodesics ...... 125

13 Floer homology 131 13.1 Near the entrance ...... 131 13.2 Morse homology in finite ...... 133 13.3 Floer homology ...... 137 13.4 An application to geodesics ...... 141 13.5 Towards the exit ...... 143

14 Non-Hamiltonian diffeomorphisms 147 14.1 The flux ...... 147 14.2 The flux conjecture ...... 150 14.3 Links to “hard” symplectic topology ...... 152 14.4 in Hofer’s geometry ...... 153 Chapter 1

Introducing the group

In this chapter we sum up some well known preliminary facts about the group of Hamiltonian diffeomorphisms.

1.1 The origins of Hamiltonian diffeomor- phisms

Consider the motion of a mass 1 particle in Rn(q), where q denotes the coordinate on Rn, in the presence of a potential force Φ(q, t) = ∂U − ∂q (q, t). Newton’s second law yieldsq ¨ = Φ(q, t). In general there is no chance to solve this equation explicitly except for some rare exam- ples. However it is possible to understand some qualitative properties. Let us make a little trick. Introduce an auxiliary variable p =q ˙, p2 and consider the function F (p, q, t) = 2 + U(q, t). The function F represents the total energy of the particle (the sum of the kinetic and the potential energies). With this notation the Newton equation above can be rewritten as follows:

 ∂F  p˙ = − ∂q (p, q, t)

 ∂F q˙ = ∂p (p, q, t). This system of the first order differential equations is called Hamil- tonian system. It should be considered as a differential equation in the

1 2 CHAPTER 1. INTRODUCING THE GROUP

2n-dimensional space R2n endowed with coordinates p and q. The first step of any qualitative study is to forget the explicit form of the ob- ject you are interested in. With this ideology at hand let us leave aside the expression for the function F above and focus on properties of general Hamiltonian systems associated with more or less arbitrary smooth energy functions F (p, q, t). “More or less” means that we im- pose certain assumptions on the behaviour of F at infinity in order to guarantee that the solutions of the Hamiltonian system exist for all 2n 2n t ∈ R. So choose such an F and consider the flow ft : R → R which takes any initial condition (p(0), q(0)) to the corresponding solution (p(t), q(t)) at time t. Informally, we call diffeomorphisms ft arising in this way mechanical motions . The essence of our trick is in the fact that these diffeomorphisms of the 2n-dimensional space R2n have the following remarkable geometric properties which cannot be seen in the original configuration space Rn. Theorem 1.1.A. (Liouville Theorem) Mechanical motions preserve the Vol = dp1 ∧ dq1 ∧ ... ∧ dpn ∧ dqn.

Theorem 1.1.B. (A more delicate variant of 1.1.A) Mechanical mo- tions preserve the 2-form ω = dp1 ∧ dq1 + ... + dpn ∧ dqn.

ωn Note that Vol = n! and hence 1.1.B implies 1.1.A. Further, for n = 1, Theorems 1.1.A and 1.1.B are equivalent. Theorem 1.1.B is a simple consequence of the fact that mechanical motions come from Hamiltonian system. We leave the proof till the next section. Both results above go back to the past. The preservation of vol- ume by mechanical motions attracted a lot of attention already more than a century ago. It served as the main stimulating force for cre- ation of ergodic theory, nowadays a well established mathematical dis- cipline which studies various recurrence properties of measure pre- serving transformations. However the significance of the role played by the invariant 2-form ω has been noticed relatively recently. As far as I know it was V.I. Arnold who first pointed it out explicitly in the 1960s. An attempt to understand the difference between mechanical motions and general volume preserving diffeomorphisms gave rise to the field of symplectic topology which investigates surprising rigidity 1.1. THE ORIGINS OF HAMILTONIAN 3

phenomena appearing in the theory of symplectic and their . Here is an example of such a phenomenon which was found by J.-C. Sikorav in [S1]. Let B2(r) ⊂ R2 be the Euclidean disc of radius r bounded by the circle S1(r) . Consider a

1 1 2 2 2n LR = S (R) × ... × S (R) ⊂ R (p1, q1) × ... × R (pn, qn) = R (p, q)

2 2n−2 and a cylinder Cr = B (r) × R . Theorem 1.1.C. (A non-squeezing property) There is no mechanical motion taking LR into Cr provided R > r. We will prove this statement in a more general context in section 3.2.E below. Note that for n = 1 the result is obvious. Indeed, the area bounded by S1(R) is greater than the area of B2(r). Thus one cannot take S1(R) into B2(r) by an area preserving transformation. However, if n ≥ 2 then LR is a of codimension n ≥ 2, while Cr has infinite volume. So there is no visible reason why the statement should hold true. Moreover, it is certainly wrong in the volume-preserving category! Exercise. Find a linear volume preserving transformation R2n → R2n taking LR into Cr for arbitrary positive r and R. In what follows we shall consider the evolution of a mechanical system as a curve in the group of all mechanical motions and study this curve by geometric tools. To make things work we are forced to restrict the class of mechanical systems we deal with. For instance, unbounded p2 Hamiltonians such as F (p, q, t) = 2 + U(q, t) above are too compli- cated to be included into such a framework. In fact we will always assume that Hamiltonian functions (and therefore the corresponding mechanical motions) are compactly supported. In other words all the action takes place in a bounded part of the space we live in. In this chapter we introduce on symplec- tic manifolds which is a natural generalization of the model described above. In this context a Hamiltonian diffeomorphism is just a mechani- cal motion generated by a compactly supported Hamiltonian function. 4 CHAPTER 1. INTRODUCING THE GROUP 1.2 Flows and paths of diffeomorphisms

To start with, we explain the relation between flows and differential equations, and give a geometric interpretation of flows as paths of diffeomorphisms. In view of future applications, we will assume that all objects we consider are compactly supported. However, basic concepts described below can be extended in a straightforward way to more general situations. Consider a smooth manifold M without boundary. For a diffeo- φ : M → M define its support supp(φ) as the closure of all x ∈ M such that φx 6= x. Denote by Diffc(M) the group of all diffeomorphisms with compact support. Let I ⊂ R be an interval. 1 A path of diffeomorphisms is a

c f : I → Diff (M), t → ft

with the following properties:

• the map M × I → M taking (x, t) to ftx is smooth;

• there exists a compact K of M which contains suppft for all t ∈ I.

We often denote such a path by {ft}. Note that on closed manifolds the second condition is automatically satisfied. Every path of diffeomorphisms generates a family of vector fields ξt, t ∈ I on M as follows: d f x = ξ (f x). (1.2.A) dt t t t

Note that this family is smooth and compactly supported: ξt(x) = 0 for all x ∈ M \ K. Such a family is called a compactly supported time-dependent vector field on M. The correspondence above is not injective. Indeed, every path of the form {ftg} where g is an arbitrary element of Diffc(M) generates the same time-dependent vector field ξ. However, for every point s ∈ I there exists unique path {ft} which

1We define an interval as a connected subset of R with non-empty interior. 1.3. CLASSICAL MECHANICS 5

generates ξ such that fs equals the identity map 1l.This path is defined as the unique solution of 1.2.A which is considered now as an ordinary differential equation with the initial condition fs = 1l. Assume that 0 ∈ I and take s = 0. The path {ft} constructed above with f0 = 1l is called a flow of the time-dependent vector field ξ. So flows are just paths {ft} satisfying f0 = 1l.

1.3 The mathematical model of classical mechanics

The role of the in classical mechanics is played by a symplectic manifold (M 2n, Ω). Here M is a connected manifold with- out boundary of even 2n, and Ω is a closed differential 2-form on M. The form Ω is assumed to be non-degenerate. This means that its top power Ωn does not vanish at any point. The form 1 n Vol = n! Ω is called the canonical volume form on (M, Ω). It is useful to keep in mind two elementary examples of symplectic manifolds: an orientable endowed with an area form, and the linear space 2n n R (p1, ..., pn, q1, ..., qn) with the form ω = Σj=1dpj ∧ dqj. The sec- ond example is of great importance in view of the classical Darboux theorem [MS]. It states that locally every symplectic manifold looks like (R2n, ω). In other words, near each point of M one can choose local coordinates (p, q) such that in these coordinates Ω is written as n Σj=1dpj ∧ dqj. We call (p, q) canonical local coordinates. Let F be a smooth function on M. A vector field ξ on M is called the Hamiltonian vector field of F if it satisfies the pointwise linear algebraic equation iξΩ = −dF . An elementary argument from linear algebra (based on the non-degeneracy of Ω) shows that ξ always exists and is unique [MS]. Sometimes ξ is denoted by sgradF (skew gradient of F ).

Exercise 1.3.A. Show that in canonical local coordinates (p, q) on ∂F ∂F M one has sgradF = (− ∂q , ∂p ). Exercise 1.3.B. Let φ : M → M be a symplectic diffeomorphism 6 CHAPTER 1. INTRODUCING THE GROUP

(that is φ∗Ω = Ω). Show that

−1 sgrad(F ◦ φ ) = φ∗sgradF

for every function F on M. This property, of course, reflects the fact that the operation sgrad is defined in the coordinate-free manner. In classical mechanics energy determines the evolution. The en- ergy is a family of functions Ft on M which depends on the additional time variable t. The time t is defined on some interval I. Equivalently, one can consider the energy as a single function F on M × I. We shall use both points of view throughout the book, and keep the no- tation Ft(x) = F (x, t). Traditionally, F is called a (time-dependent) Hamiltonian function. The evolution of the system is described by the Hamiltonian equa- tion x˙ = sgradFt(x). In local (p, q) on M the Hamiltonian equation has the familiar form (cf. 1.3.A)

 ∂F  p˙ = − ∂q (p, q, t)

 ∂F q˙ = ∂p (p, q, t). Let us introduce a linear function space A = A(M) which plays an important role below. When M is closed, define A(M) as the space of all smooth functions on M with zero mean with respect to the canonical volume form. When M is open, A(M) consists of all smooth functions with compact support. Definition. Let I ⊂ R be an interval. A (time-dependent) Hamilto- nian function F on M × I is called normalized if Ft belongs to A for all t. In the case when M is open we require in addition that there exists a compact subset of M which contains the supports of all the functions Ft, t ∈ I simultaneously. In what follows we consider normalized Hamiltonians only. Let me give some arguments in favor of this decision. First of all, on open manifolds one must impose some assumptions on the behaviour of Hamiltonian functions at infinity. Indeed, other- wise the solutions of the Hamiltonian equation may blow up at finite 1.4. THE GROUP OF HAMILTONIAN DIFFEOMORPHISMS 7

time and thus the Hamiltonian flow may be ill defined. An important feature of the definition above is that the time dependent Hamilto- nian vector field sgradFt of a normalized Hamiltonian function F is compactly supported. Thus when I contains 0 such a field generates a compactly supported flow which fits into the setting of the previous section. Second, both on open and closed manifolds, the mapping which takes a function from A to its Hamiltonian vector field is injective. In- deed, a Hamiltonian vector field determines the corresponding Hamil- tonian function uniquely up to an additive constant. Clearly, our normalization forbids to add constants! This property of normalized Hamiltonian functions will be useful for our further considerations.

1.4 The group of Hamiltonian diffeomor- phisms

Let F : M × I → R be a normalized time dependent Hamiltonian function. Assume that I contains zero. Consider the flow {ft} of the time-dependent vector field sgradFt. We will say that {ft} is the Hamiltonian flow generated by F . Every individual diffeomorphism fa, a ∈ I of this flow is called a Hamiltonian diffeomorphism. Of course, the definition implies that Hamiltonian diffeomorphisms are compactly supported.

Exercise 1.4.A (reparametrization of flows) Let {ft}, t ∈ [0; a] be a Hamiltonian flow generated by a normalized Hamiltonian F (x, t). Show that {fat}, t ∈ [0; 1] is again a Hamiltonian flow generated by aF (x, at). Therefore every Hamiltonian diffeomorphism is in fact a time-one map of some Hamiltonian flow. More generally, show that for every smooth function b(t) with b(0) = 0 the flow {fb(t)} is a Hamil- db tonian flow whose normalized Hamiltonian equals dt (t)F (x, b(t)). The crucial property of Hamiltonian diffeomorphisms is that they preserve the symplectic form Ω. Indeed let ξ be the Hamiltonian vec- tor field of a function F on M. All we have to show is that the Lie LξΩ vanishes. This can be seen from the following calcula- 8 CHAPTER 1. INTRODUCING THE GROUP tion: LξΩ = iξdΩ + d(iξΩ) = −ddF = 0. Denote by Ham(M, Ω) the set of all Hamiltonian diffeomorphisms. A path of diffeomorphisms with values in Ham(M, Ω) is called a Hamiltonian path . It sounds natural to call a flow with values in Ham(M, Ω) a Hamiltonian flow. However, after a little thinking, we realize that we are in trouble. Indeed, Hamiltonian flows were already defined above in a different way. A priori it is not at all clear why the vector field corresponding to a flow of Hamiltonian diffeomorphisms is a Hamiltonian vector field! Fortunately, this is true. This extremely important fact was established by Banyaga in [B1]. Let us formulate it precisely.

Proposition 1.4.B. For every Hamiltonian path {ft}, t ∈ I there exists a (time-dependent) normalized Hamiltonian function F : M × I → R such that d f x = sgradF (f x) dt t t t for all x ∈ M and t ∈ I.

The function F is called the normalized Hamiltonian function of {ft}. Let us discuss this result. First assume that the manifold M satis- fies the following topological condition: its first de Rham cohomology 1 with compact supports vanishes, that is Hcomp(M, R) = 0. To have an example in mind, think of the 2- or the linear space. In this case the proposition above can be proved very easily. Denote by ξt the vector field generated by ft. Since Hamiltonian diffeomorphisms

preserve the symplectic form Ω we have Lξt Ω = 0. Hence diξt Ω = 0

so iξt Ω is a closed form. In view of our topological condition the form

iξt Ω is exact. Therefore there exists unique smooth family of functions

Ft(x) ∈ A such that −dFt = iξt Ω. It follows that F (x, t) is the nor- malized Hamiltonian of {ft}. This completes the proof. However when 1 Hcomp(M, R) 6= 0, we do not know a priori that closed forms iξt Ω are exact. In this case one should use in addition that each individual ft is Hamiltonian. This requires some new ideas, see [B1],[MS]. Remark 1.4.C. Let Symp(M, Ω) be the group of all compactly sup- ported diffeomorphisms f of M which preserve Ω, that is f ∗Ω = 1.4. THE GROUP OF HAMILTONIAN DIFFEOMORPHISMS 9

Ω. Such diffeomorphisms are called . Denote by

Symp0(M, Ω) the path connected component of the identity in Symp(M, Ω). By definition, it contains those symplectomorphisms f which can be joined with the identity by a path of symplectomor- phisms. Arguing exactly as in the proof of 1.4.B above we see that 1 such a path is Hamiltonian provided Hcomp(M, R) = 0. Therefore in this case

Symp0(M, Ω) = Ham(M, Ω). 1 When Hcomp(M, R) 6= 0 this equality is not valid anymore. For in- stance, consider a 2-dimensional torus T 2 = R2(p, q)/Z2 endowed with the area form Ω = dp ∧ dq. A translation (p, q) → (p + a, q) obviously 2 lies in Symp0(T , Ω). However, for non- a one can show that it is not Hamiltonian. On the other hand the difference between Ham and Symp0 is not “too big” and it can be described in a rather simple way, see 14.1 below. The next proposition contains a remarkable elementary formula which will be used below on many occasions. Proposition 1.4.D. (Hamiltonian of the product) Consider two Hamil- tonian paths {ft} and {gt}. Let F and G be their normalized Hamil- tonian functions. Then the product path ht = ftgt is a Hamiltonian path generated by the normalized Hamiltonian function −1 H(x, t) = F (x, t) + G(ft x, t). Let us prove this formula. We are given that d d f x = sgradF and g x = sgradG . dt t t dt t t Thus d (f g )x = sgradF + f sgradG . dt t t t t∗ t −1 The second summand in the right hand side equals sgrad(G ◦ ft ) in view of Exercise 1.3.B above. We conclude that d h x = sgrad(F + G ◦ f −1) = sgradH . dt t t t t The formula is proved. Now we are ready to justify the title of this section. 10 CHAPTER 1. INTRODUCING THE GROUP

Proposition 1.4.E. The set of Hamiltonian diffeomorphisms is a group with respect to composition. Indeed, take two Hamiltonian diffeomorphisms f and g. In view of 1.4.A we can write f = f1 and g = g1 for some Hamiltonian flows {ft}, {gt} defined for t ∈ [0; 1]. Proposition 1.4.D yields that that the path {ftgt} is a Hamiltonian flow. Thus its time-one map fg is a Hamiltonian diffeomorphism. So the set Ham(M, Ω) is closed under composition of diffeomorphisms. It remains to verify that f −1 is a Hamiltonian diffeomorphism. This follows from the next exercise. −1 Exercise. Show that the path {ft } is a Hamiltonian flow generated −1 by Hamiltonian −F (ftx, t). Hint: differentiate the identity ft ◦ft = 1l with respect to t and argue as in the proof of 1.4.D above. Remark 1.4.F. In differential geometry one usually deals with groups of transformations which preserve certain structures on a manifold (such as the group Symp(M, Ω) of all symplectomorphisms.) The group of Hamiltonian diffeomorphisms does not have such a friendly descrip- tion (Hamiltonian diffeomorphisms are not defined as morphisms in certain natural category). This leads to very surprising complications. For instance the next question (known as the Flux conjecture, see Chapter 14) is still open for most symplectic manifolds M. Suppose that M is closed. Assume that some sequence of Hamiltonian dif- feomorphisms C∞-converges to a symplectic diffeomorphism f. Is f Hamiltonian? It is extremely useful to think about Ham(M, Ω) in terms of theory. This point of view is fundamental for a development of geometric intuition which is necessary for our purposes. So let us work out this language. We will consider Ham(M, Ω) as a Lie subgroup of the group of all diffeomorphisms of M. Thus the 2 of Ham(M, Ω) is just the algebra of all vector fields ξ on M of the form ¯ d ¯ ξ(x) = ¯ ftx, dt t=0 2As a vector space, the Lie algebra is by definition the tangent space to the group at the identity. The tangent spaces to the group at all other points are identified with the Lie algebra with the help of right shifts of the group. 1.4. THE GROUP OF HAMILTONIAN DIFFEOMORPHISMS 11

where {ft} is a smooth path on Ham(M, Ω) with f0 = 1l. Every such field is Hamiltonian. Indeed, ξ = sgradF0(x), where F (x, t) is the (unique!) normalized Hamiltonian function generating the path, and F0(x) = F (x, 0). Note that F0 ∈ A. Vice versa, for every function F ∈ A the vector field sgradF is by definition the derivative at t = 0 of the corresponding Hamiltonian flow. We conclude that the Lie algebra of Ham(M, Ω) can be identified with A. Exercise 1.4.G. Show that with this language the tangent vector to a Hamiltonian path {ft} at t = s is the function Fs ∈ A. Hint: The identification of tangent spaces to the group goes via right shifts (see the footnote). Thus the tangent vector in question is identified with −1 the tangent vector to the path {ftfs } at t = s. The next important notion is the adjoint action of the Lie group on its Lie algebra. Recall that this operation is defined as follows. Pick an element f of the group Ham(M, Ω) and an element G from the Lie algebra A. Let {gt}, g0 = 1l be a path on the group which is tangent to G. In our situation the tangency condition means of course that the normalized Hamiltonian function of the flow {gt} at time t = 0 equals G. By definition, ¯ d ¯ −1 Adf G = ¯ fgtf . dt t=0 Differentiating, we get that the vector field on the right hand side −1 equals f∗sgradG, but this is precisely sgrad(G◦f ) in view of Exercise 1.3.B. Returning to our identification we get that

−1 Adf G = G ◦ f .

Thus the adjoint action of Ham(M, Ω) on A is just the usual action of diffeomorphisms on functions. Finally let us discuss the Lie bracket on A. Pick two elements F,G ∈ A, and let {ft}, f0 = 1l be a Hamiltonian path tangent to F at 0. The Lie bracket {F,G} of F and G is called the and is defined as follows: ¯ d ¯ {F,G} = ¯ Adft G. dt t=0 12 CHAPTER 1. INTRODUCING THE GROUP

Calculating the expression on the right hand side we get that

{F,G} = −dG(sgradF ) = Ω(sgradG, sgradF ).

Let us mention that in terms of vector fields, the Lie bracket coincides with the usual commutator up to the sign. The reader is invited to check that [sgradF, sgradG] = −sgrad{F,G},

where the commutator [X,Y ] of two vector fields is defined by L[X,Y ] = LX LY −LY LX . WARNING: different authors may use different signs in the definitions of the following notions which play an important role in this book: Hamiltonian vector field , Poisson bracket, commutator of vector fields, and curvature of a connection.

Example 1.4.H. Consider the unit sphere S2 inside the R3. Let Ω be the induced area form on the sphere. The group SO(3) acts on S2 by area preserving diffeomorphisms. Since SO(3) 2 is path connected it is contained in Symp0(S ). Applying 1.4.C we 2 2 get that Symp0(S ) = Ham(S ), and thus SO(3) is a subgroup of Ham(S2). In particular every element of the Lie algebra so(3) can be uniquely represented as a normalized Hamiltonian function on S2. We wish to describe this correspondence precisely. Identify so(3) with R3 as follows. Every vector a ∈ R3 is considered as a skew-symmetric transformation x → [x, a] of the space, where the brackets stand for the standard vector product. Identify the tangent to S2 at a point x with the orthogonal complement to x in the space. By tauto- logical reasons, the Hamiltonian vector field v of the flow x → exp(ta)x on the sphere is given by v(x) = [x, a]. We claim that the correspond- ing normalized Hamiltonian is the height function F (x) = (a, x). First of all, the reflection over the orthogonal complement to a takes F to −F , thus F has zero mean. Further, note that Ω(ξ, η) = (η, [x, ξ]) for 2 0 2 ξ, η ∈ TxS . Denote by a the orthogonal projection of a to TxS . Thus 0 2 Ω(ξ, v(x)) = ([x, a ], [x, ξ]) for every ξ ∈ TxS . Since the vector prod- 2 uct with x is an orthogonal transformation of TxS , the last expression equals (a0, ξ). But this is precisely dF (ξ). The claim follows. 1.5. ALGEBRAIC PROPERTIES OF HAM(M, Ω) 13 1.5 Algebraic properties of Ham(M, Ω)

Algebraic properties of the group of Hamiltonian diffeomorphisms were studied by A. Banyaga in [B1],[B2]. In particular, he proved the fol- lowing striking result. Recall, that a group D is called simple if every is trivial, that is equals either {1l} or D itself. Theorem 1.5.A. Let (M, Ω) be a closed symplectic manifold. Then the group Ham(M, Ω) is simple. There exists a version of this statement for open M as well. Note that an is simple if and only if every ele- ment generates the whole group (and so it must be a finite whose order is a prime number). Thus intuitively speaking gen- eral simple groups are far from being abelian. Below we present an elementary statement which clarifies this principle for the group of Hamiltonian diffeomorphisms. We will use it in the next chapter. Proposition 1.5.B. Let (M, Ω) be a symplectic manifold and let U ⊂ M be a non-empty open subset. There exist f, g ∈ Ham(M, Ω) such that supp(f), supp(g) ⊂ U and fg 6= gf. The proof is based on the following fact.

Proposition 1.5.C. Let {ft} and {gt} be the Hamiltonian flows gen- erated by time independent normalized Hamiltonian functions F and G respectively. If ftgt = gtft for all t then {F,G} = 0.

Proof: According to 1.4.D the corresponding Hamiltonians for ftgt and gtft are respectively

−1 −1 F (x) + G(ft (x)) and G(x) + F (gt (x)).

But since the flows are the same for both we have that

−1 −1 F (x) + G(ft (x)) = G(x) + F (gt (x)) for all t. Differentiating with respect to t we get

dG(−sgradF ) = dF (−sgradG) 14 CHAPTER 1. INTRODUCING THE GROUP which yields {F,G} = {G, F }. Hence, from the anti-commutativity of the Lie bracket, {F,G} = 0. 2

Proof of 1.5.B: Choose a point x ∈ U and tangent vectors ξ, η ∈ TxU such that Ω(ξ, η) 6= 0. Now choose germs of functions F and G (see exercise below) satisfying sgradF (x) = ξ, sgradG(x) = η. Extend these functions by 0 outside U. If M is open we are done. If M is closed, add a constant to guarantee that F and G have zero mean. So now the functions F,G belong to A. Further, they are constant outside U, and so the corresponding Hamiltonian diffeomorphisms ft and gt are supported in U. Since {F,G} 6= 0 we see that for some t, the diffeomorphisms ft and gt do not commute. 2

Exercise. Use local canonical coordinates near x to show that F and G in the proof of the corollary do exist. We complete this section with the following result due to A. Banyaga (along the lines of [B2]).

Theorem 1.5.D. Let (M1, Ω1) and (M2, Ω2) be two closed symplectic manifolds such that their groups of Hamiltonian diffeomorphisms are isomorphic. Then the manifolds are conformally symplectomorphic: there exists a diffeomorphism f : M1 → M2 and a number c 6= 0 such ∗ that f Ω2 = cΩ1.

In other words, the of the group of Hamiltonian diffeomorphisms determines the symplectic manifold up to a factor. Chapter 2

Introducing the geometry

In the present chapter we discuss biinvariant Finsler metrics on the group of Hamiltonian diffeomorphisms and define Hofer’s geometry.

2.1 A variational problem

What is the minimal amount of energy needed in order to generate a given Hamiltonian diffeomorphism φ? This natural question can be formalized as follows. Consider all possible Hamiltonian flows {ft}, t ∈ [0; 1] such that f0 = 1l and f1 = φ. For each flow take its unique normalized Hamiltonian function Ft(x) and “measure its magnitude”. Then minimize the result of the measurement over all such flows. It remains to explain what we mean by the “magnitude”. Recall that for every t the function Ft is an element of the Lie algebra A. Choose any norm || || on A in coordinate-free way, that is we require that

||H ◦ ψ−1|| = ||H|| for all H ∈ A, ψ ∈ Ham(M, Ω). (2.1.A)

R 1 Now define the magnitude as 0 ||Ft||dt. Bringing together all ingre- dients of this procedure we end up with the variational problem

Z 1 inf ||Ft||dt, (2.1.B) 0

where φ is given, and the infimum is taken over all flows {ft} as above.

15 16 CHAPTER 2. INTRODUCING THE GEOMETRY 2.2 Biinvariant geometries on Ham(M, Ω)

It turns out that the variational problem above can be reformulated in purely geometrical terms. To make this transparent I wish to recall the notion of a Finsler structure on a manifold. We say that a manifold Z is endowed with a Finsler structure if its tangent spaces TzZ are equipped with a norm which depends smoothly on the points z ∈ Z. Of course, Riemannian structures form a particular case of this notion. In general however the norms above may not come from a scalar product. Given a Finsler structure, one defines the length of a curve exactly as in the Riemannian case by

Z b length{z(t)}t∈[a;b] = norm(z ˙(t))dt. a Further, one introduces the distance between two points z and z0 in Z as the infimum of lengths of all curves joining z and z0. Let us return to the situation described in the previous section. Since all tangent spaces to the group Ham(M, Ω) are identified with A (see section 1.4 above), every choice of the norm || || on A leads to a Finsler structure on the group. Thus one can define the length of a Hamiltonian path, and the distance between two Hamiltonian diffeo- morphisms. In particular, the length of a Hamiltonian path {ft}, t ∈ [a; b] with the normalized Hamiltonian F is given by

Z b length{ft} = ||Ft||dt. a The distance between two Hamiltonian diffeomorphisms φ and ψ is defined by ρ(φ, ψ) = inf length{ft}, where the infimum is taken over all Hamiltonian paths {ft}, t ∈ [a; b] with fa = φ and fb = ψ. Of course the length of a path does not depend on the parametrization, thus in the definition of the distance above one can take a = 0 and b = 1. With this language, the solution of the variational problem 2.1.B above is nothing else but the distance ρ(1l, φ)! 2.2. BIINVARIANT GEOMETRIES ON HAM(M, Ω) 17

The following properties of ρ are easily verified and are left as an exercise:

• ρ(φ, ψ) = ρ(ψ, φ);

• the triangle inequality : ρ(φ, ψ) + ρ(ψ, θ) ≥ ρ(φ, θ);

• ρ(φ, ψ) ≥ 0.

Recall now condition 2.1.A imposed on the norm || || in section 2.1 above. In geometric language this condition means that the norm is invariant under the adjoint action of the group on the Lie algebra (see 1.4 above). In what follows we deal with such norms only.

Exercise. Show that 2.1.A yields that the function ρ is biinvariant: 1

ρ(φ, ψ) = ρ(φθ, ψθ) = ρ(θφ, θψ) for all φ, ψ, θ ∈ Ham(M, Ω).

It would be more honest to call the function ρ a pseudo-distance. Indeed, as we have seen above it satisfies all the axioms of a distance function on a metric space except possibly the non-degeneracy, that is

ρ(φ, ψ) > 0 for φ 6= ψ. (2.2.A)

It is rather non-trivial to check the non-degeneracy even in finite dimensional geometry. The argument in this case uses local compact- ness of manifolds. In our situation the group Ham(M, Ω) is infinite dimensional and has no compactness properties. Thus a priori there is no reason for 2.2.A to be true. It turns out that the non-degeneracy of ρ is very sensitive to the choice of the norm || ||.

1Without assumption 2.1.A we get only right-invariant ρ’s, that is ρ(φ, ψ) = ρ(φθ, ψθ). Such metrics play an important role in hydrodynamics, see [AK]. 18 CHAPTER 2. INTRODUCING THE GEOMETRY

2.3 The choice of the norm: Lp vs. L∞ Among the norms on A satisfying the invariancy assumption 2.1.A there is a very natural class which includes Lp norms, p = 1, 2, 3, ... Z ¡ ¢ 1 p p kHkp = |H| Vol , M

and the L∞-norm

kHk∞ = max H − min H.

Denote by ρp and ρ∞ the corresponding pseudo-distances.

Theorem 2.3.A. The pseudo-distance ρp is degenerate for all finite p = 1, 2, .... Moreover, if the manifold is closed then all such ρp vanish identically.

This result was established in [EP] . The proof is presented below in this chapter (see also books [HZ], [MS], [AK]). The next theorem shows a striking contrast between the Lp and the L∞ cases.

Theorem 2.3.B. The pseudo - distance ρ∞ is non-degenerate.

This theorem 2 was discovered and proved by Hofer in [H1] for the case M = R2n with the use of infinite dimensional variational methods. In [V1] Viterbo derived it for the case M = R2n from his theory of gen- erating functions. Both Hofer and Viterbo obtained their results as an answer to a stimulating question posed by Eliashberg in a private discussion. In [P1] the statement was extended to a wide class of sym- plectic manifolds with “nice” behaviour at infinity, and in particular to all closed symplectic manifolds such that the cohomology class of the symplectic form is rational. The approach of [P1] is based on Gro- mov’s theory of pseudo-holomorphic curves. Finally [LM1] Lalonde and McDuff proved 2.3.B in full generality using Gromov’s theory. At present some other proofs of various particular cases of this theorem are available, see for instance [Ch],[O3],[Sch3]. Hofer’s original proof

2The historical digression below reflects my own understanding of the situation. I admit that other participants of this development may see it differently. 2.4. THE CONCEPT OF DISPLACEMENT ENERGY 19

is presented in great details in the book [HZ]. The argument due to Lalonde and McDuff is outlined in the book [MS] and in the survey [L]. Below we give a different proof for the case of M = R2n which follows [P1]. All known proofs are based on “hard” methods. 3

2.4 The concept of displacement energy

Which invariant norms || || on A (that is norms satisfying 2.1.A) lead to non-degenerate distance functions ρ ? Below we describe a very useful reformulation of this question which will eventually enable us to prove Theorem 2.3.A. It is based on the beautiful concept of displacement energy which was introduced by Hofer in [H1]. Let ρ be a biinvariant pseudo-distance on Ham(M, Ω), and let A be a bounded subset of M. Definition. The displacement energy of A is given by

e(A) = inf{ρ(1l, f)| f ∈ Ham(M, Ω), f(A) ∩ A = ∅}.

The set of such f may be empty. In what follows we use the convention that the infimum of the empty set equals +∞. If e(A) 6= 0 we will say that A has positive displacement energy. Let us mention two obvious but important properties of e(A). First of all, e is a monotone function of : if A ⊂ B then e(A) ≤ e(B). Second, e is an invariant, that is e(A) = e(f(A)) for every Hamiltonian diffeomorphism f of M. We leave the proofs to the reader. Example. Consider (R2, ω) and take an open square A whose edges have length u and are parallel to the coordinate axes. Let us esti- mate the displacement energy of A with respect to the distance ρ∞ . Consider the Hamiltonian function H(p, q) = up. The corresponding Hamiltonian system is ½ q˙ = u p˙ = 0

3Moreover, to my taste all the proofs are far from being transparent. The argument presented in Chapter 3 below is not an exception. I strongly believe that in the future one will find the explanation of this fundamental result. 20 CHAPTER 2. INTRODUCING THE GEOMETRY

Therefore its time-1-map h sends (p, q) to (p, q + u). Note that all the motion of the square takes place in the rectangle K = Closure(A ∪ h(A)). Consider a cut off F of the Hamiltonian H outside a small neighbourhood of K. 4 Note that F (in contrast to H) is a normalized Hamiltonian function. Since F = H on K, the time one map f of the Hamiltonian flow generated by F still displaces the square A. We can always perform the cutting off in such a way that the L∞-norm of F is arbitrary close to the oscillation of H on K. This oscillation equals

max H − min H = u2 − 0 = u2. K K Thus we get that e(A) ≤ u2 = Area(A). Note that the square is symplectomorphic to the disc of the same area in R2 (this is not true in higher dimensions!). Thus we proved that e(B2(r)) ≤ πr2. A deep result due to Hofer ([H1]) states that actually there is equality in all dimensions, in other words e(B2n(r)) = πr2. See [LM1] for a generalization to arbitrary symplectic manifolds. We will prove a lower bound for e(B2n(r)) in the next chapter. In general, if the displacement energy of all non-empty open sub- sets with respect to some biinvariant pseudo-metric ρ is positive then ρ is non-degenerate. Indeed, each f ∈ Ham(M, Ω) such that f 6= 1l must displace some small A ⊂ M. Thus we get that ρ(1l, f) ≥ e(A) > 0. In fact, the converse is also true. Theorem 2.4.A. ([EP]) If ρ is non-degenerate then e(A) > 0 for every non-empty open subset A. In order to prove this theorem we need an auxiliary lemma. Lemma 2.4.B. Let A ⊂ M be a non-empty open subset. For all φ, ψ ∈ Ham(M, Ω) with supp(φ) ⊂ A and supp(ψ) ⊂ A, we have 1 e(A) ≥ 4 ρ(1l, [φ, ψ]). 4Let Y be a closed subset of a manifold Z, and let H be a smooth function defined in a neighbourhood V of Y . By a cut off F of H outside a small neigbour- hood of Y we mean the following. Choose a neighbourhood W of Y whose closure is contained in V . Take a smooth function a : Z → [0; 1] which equals 1 on W and vanishes outside V . Define now F by aH on V and by 0 outside V . 2.4. THE CONCEPT OF DISPLACEMENT ENERGY 21

Here [φ, ψ] stands for the commutator ψ−1φ−1ψφ. The theorem immediately follows from the lemma in view of Proposition 1.5.B of the previous chapter. Proof of 2.4.A: By 1.5.B there exist φ, ψ supported in A with [φ, ψ] 6= 1 1l. Since ρ is non-degenerate this implies that e(A) ≥ 4 ρ(1l, [φ, ψ]) > 0. 2

Proof of 2.4.B: Assume that there exists h ∈ Ham(M, Ω) such that h(A) ∩ A = ∅ (if such an h does not exist we are done because e(A) = +∞). Set θ = φh−1φ−1h = [h, φ−1]. If x ∈ A then h(x) ∈/ A. Since φ = 1loutside A we get φ−1h(x) = h(x). So we see that h−1φ−1h(x) = x and −1 −1 therefore θ|A = φ|A. Now supp(ψ) ⊂ A and therefore φ ψφ = θ ψθ which implies that [φ, ψ] = [θ, ψ]. Note that

ρ(1l, [θ, ψ]) = ρ(1l, ψ−1θ−1ψθ) = ρ(θ−1, ψ−1θ−1ψ) ≤ ρ(1l, θ−1) + ρ(1l, ψ−1θ−1ψ) = 2ρ(1l, θ).

Here we have used the bi-invariance of ρ and the triangle inequality. Analogously, ρ(1l, θ) = ρ(1l, [h, φ−1]) ≤ 2ρ(1l, h). Putting these two inequalities together we get

ρ(1l, [φ, ψ]) = ρ(1l, [θ, ψ]) ≤ 2ρ(1l, θ) ≤ 4ρ(1l, h).

Since this holds for all h ∈ Ham(M, Ω) with h(A) ∩ A = ∅ we obtain, by taking the infimum, that 4e(A) ≥ ρ(1l, [φ, ψ]). 2

Recall that Theorem 2.3.A states that for p < ∞ the Lp-norm leads to a degenerate pseudo-distance, which in fact vanishes on closed manifolds. Now we will prove this statement and see why the argument breaks down in the L∞-case. Proof of 2.3.A: We will show that the displacement energy of a ’small’ ball vanishes. Then the degeneracy of ρp follows from 2.4.A. Let U be an open subset of M endowed with canonical coordinates (x, y). 22 CHAPTER 2. INTRODUCING THE GEOMETRY P In these coordinates the symplectic form Ω is given by dxi ∧ dyi. P 2 2 Assume without loss of generality that U contains a ball (xj +yj ) < 10. Let A ⊂ U be the ball with the same center of the radius 0.1. Consider a (partially defined) flow ht, t ∈ [0; 1] on U which is simply a shift by t along the y1 coordinate. Such a shift is generated by a (non-normalized!) Hamiltonian function H(x, y) = x1 on U. Clearly, h1(A)∩A = ∅. Denote by St the sphere ht(∂A). Consider a new (time- dependent) normalized Hamiltonian function Gt = Ft + ct where Ft is obtained from H by a cut off outside a small neighbourhood of St, and ct is a (time dependent)R constant. Here of course ct = 0 if the manifold M is open, and ct = − M FtVol if M is closed. Since for every t the function Gt coincides with H near St up to an additive constant we conclude that sgrad Gt = sgrad H near St. Hence the Hamiltonian flow {gt} of G satisfies gt(∂A) = ht(∂A) and so g1(∂A) ∩ ∂A = ∅. But this obviously implies that g1(A) ∩ A = ∅. Note now that using cutting off outside very small neighbourhoods of St we can achieve that the Lp-norm of every function Gt is arbitrary small (in contrast to the L∞-norm!). Therefore the Lp- displacement energy of A vanishes. This completes the proof of the degeneracy of ρp. Let us turn now to the second statement of the theorem, where we assume that the manifold M is closed. Set G = {g ∈ Ham(M, Ω) | ρ(1l, g) = 0}. Take f, g ∈ G. Then of course g−1 ∈ G. Further the triangle inequality yields that ρ(1l, fg) = ρ(f −1, g) ≤ ρ(1l, f) + ρ(1l, g) = 0 so G is a subgroup of Ham(M, Ω). By the bi-invariance we know that if f ∈ G and h ∈ Ham(M, Ω) then hfh−1 ∈ G, and so G is a nor- mal subgroup. In view of Banyaga’s theorem 1.5.A above the group Ham(M, Ω) is simple. Therefore either G = {1l} or G = Ham(M, Ω). We already proved that ρp is degenerate, hence G 6= {1l}. Thus G coin- cides with the whole group Ham(M, Ω). We conclude that ρp vanishes identically. 2 Exercise. Prove that the displacement energy of S2n−2 ⊂ R2n−1 ⊂ 2n R with respect to ρ∞ vanishes. On the other hand, we will see in 2.4. THE CONCEPT OF DISPLACEMENT ENERGY 23

the next chapter that there exist half-dimensional submanifolds of R2n which have positive displacement energy (cf. 1.1.C above). Open problem. Which invariant norms on A give rise to non-degene- rate distance functions ρ? Is it true that such norms are always bounded below by const|| ||∞? A difficulty here is that no classification of Ham(M, Ω)-invariant norms is known. A potential approach to this problem would be to investigate cut-offs. If cutting off decreases the norm arbitrarily then our argument above shows that ρ is degenerate. Open problem. [EP] It is quite natural to consider separately the positive and the negative parts of the metric ρ∞. More explicitly, set

Z 1 ρ+(1l, f) = inf max Ftdt, 0 and Z 1 ρ−(1l, f) = inf − min Ftdt. 0 Then clearly ρ(1l, f) ≥ ρ+(1l, f) + ρ−(1l, f). However in all examples known to me equality holds! It would be interesting to prove a general statement, or to find a counterexample. 2n Note that it follows from [V1] that on Ham(R ) the sum ρ+ + ρ− defines a biinvariant metric. As far as I know, no analogue of this statement was proved for general symplectic manifolds. Convention. Unless otherwise stated in what follows we use notation || || for the L∞-norm on A(M). We write ρ for the metric ρ∞ and call it Hofer’s metric . The quantity ρ(1l, f) is called Hofer’s norm of f. We write length{ft} for the length of a Hamiltonian path {ft} with respect to the L∞-norm (see 2.2 above). 24 CHAPTER 2. INTRODUCING THE GEOMETRY Chapter 3

Lagrangian submanifolds

The purpose of the next two chapters is to prove that Hofer’s met- ric on R2n is non-degenerate. We use an approach of [P1]. For that purpose we introduce Lagrangian submanifolds of symplectic mani- folds. Lagrangian submanifolds play a fundamental role in symplectic topology as well as in its applications to mechanics and calculus of variations. They will appear on many occasions throughout this book.

3.1 Definitions and examples

Definition. Let (M 2n, Ω) be a symplectic manifold and let L ⊂ M 1 be a submanifold. We call L Lagrangian if dim L = 2 dim M = n n 2n and Ω|TL ≡ 0. An (or ) f : L → M is called Lagrangian if f ∗Ω ≡ 0.

Let us list some important examples of Lagrangian submanifolds. 3.1.A. Curves on surfaces Let (M 2, Ω) be an oriented surface with an area form. Then every curve is Lagrangian (since the tangent space to a curve is one dimen- sional and Ω vanishes when it is evaluated on two vectors which are proportional). 3.1.B. The split torus (cf. 1.1.C above).

25 26 CHAPTER 3. LAGRANGIAN SUBMANIFOLDS

The torus S1 × ... × S1 ⊂ R2 × ... × R2 = R2n is Lagrangian (the 2-form Ω on R2n splits). 3.1.C. Graphs of closed 1-forms in cotangent bundles This example plays an important role in classical mechanics. Let N n be any manifold. Consider the M = T ∗N of N with the natural projection π : T ∗N → N, (p, q) 7→ q. Define the following 1-form λ on M which is called the Liouville form. For q ∈ ∗ ∗ N, (p, q) ∈ T N and ξ ∈ T(p,q)T N we set λ(ξ) = hp, π∗ξi where h , i is the natural pairing between T ∗N and TN. We claim that Ω = dλ is a ∗ symplectic form on T M. We use local coordinates (p1, ..., pn, q1, ..., qn) ∗ on T N and write ξ = (p ˙1, ..., p˙n, q˙1, ..., q˙n),P so π∗ξ = (q ˙1, ..., q˙n). With this notationP the pairing reads hp, π∗ξiP= piq˙i which implies that λ(ξ) = pidqi. Therefore Ω = dλ = dpi ∧ dqi, and we recognize the standard symplectic form on R2n. The claim follows. Exercise. Let α be a 1-form on N. Show that graph(α) is a Lagrangian submanifold of T ∗N if and only if α is closed. 3.1.D. Symplectomorphisms as Lagrangian submanifolds Let f :(M, Ω) → (M, Ω) be a diffeomorphism. Consider a new symplectic manifold (M × M, Ω ⊕ −Ω). We leave it as an exercise to show that graph (f) ⊂ (M × M, Ω ⊕ −Ω) is Lagrangian if and only if f is a . 3.1.E. Lagrangian suspension Let L ⊂ (M, Ω) be a Lagrangian submanifold. Consider a loop of 1 Hamiltonian diffeomorphisms {ht}, t ∈ S , h0 = h1 = 1l generated by a 1-periodic Hamiltonian function H(x, t). Proposition. Let M×T ∗S1 be a symplectic manifold with symplectic form σ = Ω+dr∧dt where M is as above and (r, t) are the coordinates on T ∗S1 = R × S1. Then 1 ∗ 1 φ : L × S → M × T S , (x, t) 7→ (ht(x), −H(ht(x), t), t) is a Lagrangian embedding. Proof: It suffices to prove that φ∗σ vanishes on pairs (ξ, ξ0) and on ∂ 0 ∂ 1 pairs (ξ, ∂t ) for ξ, ξ ∈ TL and ∂t ∈ TS . One computes that ∂ φ ξ = h ξ − hdH , h ξi , ∗ t∗ t t∗ ∂r 3.2. THE LIOUVILLE CLASS 27

∂ φ ξ0 = h ξ0 − hdH , h ξ0i , ∗ t∗ t t∗ ∂r ∗ 0 0 0 so φ σ(ξ, ξ ) = Ω(ht∗ξ, ht∗ξ ) = Ω(ξ, ξ ) = 0 since L is Lagrangian. Furthermore ∂ ³ ∂H ´ ∂ ∂ φ = sgrad H − hdH , sgrad H i + + ∗ ∂t t t t ∂t ∂r ∂t ∂H ∂ ∂ = sgrad H − + , t ∂t ∂r ∂t so that we get ∂ ∂ ∂ φ∗Ω(ξ, ) = Ω(h ξ, sgrad H ) + dr ∧ dt(−hdH , h ξi , ) ∂t t∗ t t t∗ ∂r ∂t = Ω(ht∗ξ, sgrad Ht) − hdHt, ht∗ξi

= dHt(ht∗ξ) − hdHt, ht∗ξi = 0.

2

3.2 The Liouville class of Lagrangian sub- manifolds in R2n

Let L ⊂ (R2n, dp ∧ dq) be a Lagrangian submanifold. Consider the restriction λ|TL of the Liouville form

λ = p1dq1 + ... + pndqn

to L. Clearly d(λ|TL) = Ω|TL = 0. The cohomology class λL ∈ H1(L, R) of this closed 1-form is called the Liouville class of the Lagrangian submanifold L. Similary, for a Lagrangian embedding or immersion φ : L → R2n we define the Liouville class as [φ∗λ]. Ge- ometrically, the Liouville class of a Lagrangian submanifold can be interpreted as follows. Let a ∈ H1(L) be a 1-cycle. Then we can find a 2-chain Σ in R2n with ∂Σ = a. Now Z Z

(λL, a) = λL = Ω a Σ 28 CHAPTER 3. LAGRANGIAN SUBMANIFOLDS

and it is easy to see that this is independent of the choice of Σ. In view of this formula the value (λL, a) is sometimes called the symplectic area of a. The last construction generalizes to an arbitrary Lagrangian manifold L of a symplectic manifold M and defines a natural homo- morphism H2(M,L; Z) → R. An important property of λL is that it 2n ∗ is invariant under symplectomorphisms of R i.e. f λf(L) = λL. Theorem 3.2.A. ([G1]) Assume that L ⊂ R2n is a closed Lagrangian submanifold . Then λL 6= 0. Let us emphasize that here L is embedded. For Lagrangian im- mersions this statement is in general not true. In the case n = 1 this can be seen as follows. It is clear that any closed embedded curve bounds a subset of positive area, but for example the immersed figure eight curve can bound a set with zero area. This phenomenon reflects “rigidity of Lagrangian ”. Definition. A closed Lagrangian submanifold L ⊂ (R2n, ω) is called rational if λL(H1(L; Z)) ⊂ R is a discrete subgroup. We will denote its positive generator by γ(L). Example. The split torus L = S1(r) × ... × S1(r) ⊂ R2n is rational. 1 2 2 Since each S (r) has symplectic√ area πr we have γ(L) = πr . How- 1 1 3 4 ever, the torus S (1) × S ( 2) ⊂ R is not rational.√ The symplectic areas of the two circles are respectively π and 3 4π and they generate a dense subgroup of R. Theorem 3.2.B. ([S1]) Let L ⊂ B2(r) × R2n−2 be a closed rational Lagrangian submanifold. Then γ(L) ≤ πr2. Note that the assumption that L is embedded is necessary. Figure 1 shows an immersed Lagrangian submanifold of arbitrary symplectic area.

Our next result gives a lower bound for the displacement energy e(L) of a rational Lagrangian submanifold L with respect to Hofer’s metric. 3.2. THE LIOUVILLE CLASS 29

Figure 1

Theorem 3.2.C. Let L ⊂ R2n be a closed rational Lagrangian submani- 1 fold. Then e(L) ≥ 2 γ(L). Theorems 3.2.A and 3.2.B are proved in the next chapter. The- orem 3.2.C follows from 3.2.B (see section 3.3 below). Let us derive some consequences of these results. 3.2.D. Non-degeneracy of Hofer’s metric. Theorem 3.2.C implies that Hofer’s metric on Ham(R2n, ω) is non-degenerate. Indeed, each ball 2n 2 2 2 2 2 B (r) = {p1 + ... + pn + q1 + ... + qn ≤ r } contains a rational split torus r r n r2 r2 o S1(√ ) × ... × S1(√ ) = p2 + q2 = , ..., p2 + q2 = . n n 1 1 n n n n

2n πr2 Thus e(B (r)) ≥ 2n > 0, and as it was explained in 2.4 above we get the desired non-degeneracy of ρ. This estimate is not sharp. Hofer proved [H1] that e(B2n(r)) = πr2. 3.2.E. The non-squeezing property. Note that γ(L) is a symplectic in- variant i.e. for a symplectomorphism f : R2n → R2n we have γ(f(L)) = γ(L). Thus Theorem 3.2.B implies the non-squeezing Theorem 1.1.C. Recall that it states that a split torus with large γ(L) = πR2 cannot be moved by a Hamiltonian diffeomorphism into B2(r) × R2n−2 for r < R. 3.2.F. The cylindrical symplectic capacity. Let A ⊂ R2n be a bounded subset and set

c(A) = inf{πr2|∃ g : R2n → R2nwith g(A) ⊂ B2(r) × R2n−2} , where g is a symplectomorphism. This function defined on subsets of R2n is called the cylindrical symplectic capacity. In this language 30 CHAPTER 3. LAGRANGIAN SUBMANIFOLDS

Theorem 3.2.B reads that for a closed rational Lagrangian submanifold c(L) ≥ γ(L). This capacity is a symplectic invariant and satisfies the following monotonicity property. For A ⊂ B, c(A) ≤ c(B) (compare to a similar monotonicity property of the displacement energy, see 2.4 above).

3.2.G. Some generalizations. Let (M, Ω) be a symplectic manifold. When M is open we assume that it has a ”nice” behaviour at infinity (this class includes for instance any cotangent bundle endowed with the standard symplectic structure, as well as a product of a cotan- gent bundle with any closed symplectic manifold). Take a Lagrangian submanifold L ⊂ M, and consider a homomorphism λL : π2(M,L) → R which takesR any disc Σ in M whose boundary lies on L to its sym- 2n plectic area Σ Ω. Exactly as in the case M = R we say that L is rational if the of λL is discrete, and for a rational L define γ(L) as the positive generator of the image of λL. When λL = 0 put γ(L) = +∞. Our proof of Theorem 3.2.C extends without essential changes to this more general setting (see [P1]), namely one gets that 1 1 e(L) ≥ 2 γ(L). As a corollary, we obtain the following important fact proved by Gromov in [G1]: e(L) = +∞ when λL = 0. This can be interpreted as a Lagrangian intersection property: if λL = 0 then φ(L) intersects L for every Hamiltonian diffeomorphism φ. We discuss an application of this result to Hofer’s geometry in Chapter 6 below. Let us mention also that these estimates were significantly im- proved in [Ch] with the use of Floer homology (see also [O3]). In particular it was shown that every (not necessarily rational) closed Lagrangian submanifold L ⊂ M has positive displacement energy.

3.2.H. An isoperimetric inequality. We conclude this section with the following beautiful result due to Viterbo [V2]. Let L ⊂ R2n be a closed Lagrangian submanifold. Denote by V the n-dimensional Euclidean volume of L. Then e(L) ≤ 2n(n−1)/2nnV.

The precise constant is still to be found.

1This point was missed in [P1, p. 359]. 3.3. ESTIMATING THE DISPLACEMENT ENERGY 31 3.3 Estimating the displacement energy

In this section we deduce Theorem 3.2.C from Theorem 3.2.B by an elementary geometric argument. Step 1. Let L be a closed rational Lagragian submanifold and let ht, t ∈ [0, 1] be a path of Hamiltonian diffeomorphisms such that h0 = 1l and h1(L) ∩ L = ∅. Fix ε > 0. Without loss of generality one can assume that ht = 1l for t ∈ [0, ε] and ht = h1 for t ∈ [1 − ε, 1]. Indeed, one can achieve this by a suitable reparametrization of the flow which preserves its length (use Exercise 1.4.A above). Let H(x, t) be the corresponding Hamiltonian function. Set Z 1 l = length {ht} = max Ht − min Htdt. 0 x x 1 We have to prove that l ≥ 2 γ(L). The master plan is to encode the motion of L under the flow as a closed Lagrangian submanifold in R2n+2 and then obtain the result using 3.2.B. We will make use of the Lagrangian suspension construction described before. In order to do this we need a loop of Hamiltonian diffeomorphisms. Step 2. We create a loop of Hamiltonian diffeomorphisms for t ∈ [0, 2] ½ ht for t ∈ [0, 1] gt = h2−t for t ∈ [1, 2] with corresponding Hamiltonian ½ H(x, t) for t ∈ [0, 1] G(x, t) = −H(x, 2 − t) for t ∈ [1, 2].

R 2 Exercise. Show that 0 G(gt(x), t)dt = 0 for all x. Apply now the Lagrangian suspension construction (see 3.1.E above). We get a new Lagrangian submanifold L0 ⊂ R2n ×T ∗S1 as the image of L × S1 under the mapping

(x, t) 7→ (gt(x), −G(gt(x), t), t). Note that we take here S1 = R/2Z. Define two functions

a+(t) = − min G(x, t) + ε x 32 CHAPTER 3. LAGRANGIAN SUBMANIFOLDS and a−(t) = − max G(x, t) − ε. x Clearly L0 ⊂ R2n × C ⊂ R2n × T ∗S1, where C stands for the annulus {a−(t) < r < a+(t)} (see figure 2). Step 3. We now wish to pass from R2n × T ∗S1 to R2n × R2. For this purpose we consider a special symplectic immersion θ : C → R2 (this is known as Gromov’s figure eight trick, see [G1], [AL] ). Exercise. (See figure 3) Show that there exists a symplectic immersion θ : C → R2(p, q) with the following properties: • θ takes the zero-section {r = 0} to the figure eight curve with ears of equal area. Thus the closed form θ∗pdq − rdt is exact (it is closed because θ is a symplectomorphism). • θ is an embedding outside the thin necks and glues the thin necks together. • the area of the interior ears is arbitrarily small, say ε each.

r

C

¡ ¡ ¡ ¡ ¡

¤ ¤ ¤ ¤

¥ ¥ ¥ ¥

¢ ¢ ¢ ¢ ¢

£ £ £ £ £

ε ¡ ¡ ¡ ¡ ¡

¤ ¤ ¤ ¤

¥ ¥ ¥ ¥

¢ ¢ ¢ ¢ ¢

£ £ £ £ £

¡ ¡ ¡ ¡ ¡

¤ ¤ ¤ ¤

¥ ¥ ¥ ¥

¢ ¢ ¢ ¢ ¢ ¦

£ £ £ £ £ §

¡ ¡ ¡ ¡ ¡

¤ ¤ ¤ ¤

¥ ¥ ¥ ¥

¢ ¢ ¢ ¢ ¢

£ £ £ £ £

¡ ¡ ¡ ¡ ¡

¤ ¤ ¤ ¤

¥ ¥ ¥ ¥

¢ ¢ ¢ ¢ ¢

£ £ £ £ £

¡ ¡ ¡ ¡ ¡

¤ ¤ ¤ ¤

¥ ¥ ¥ ¥

¢ ¢ ¢ ¢ ¢

£ £ £ £ £

−ε ε 2−ε

¡ ¡ ¡ ¡ 1−ε ¡ 1+ε 2 t

thin necks Figure 2

3.3. ESTIMATING THE DISPLACEMENT ENERGY 33

¢ ¢ ¢ ¢

graph of¢ a

£ £ £ £ £ -

interior ears

¢ ¢ ¢ ¢ ¢

£ £ £ £ £

¤ ¤ ¤ ¤ ¤ ¤ ¤

¥ ¥ ¥ ¥ ¥ ¥ ¥

¢ ¢ ¢ ¢ ¢

£ £ £ £

£ graph of a

¤ ¤ ¤ ¤ ¤ ¤

¤ +

¥ ¥ ¥ ¥ ¥ ¥ ¥

¢ ¢ ¢ ¢ ¢

£ £ £ £ £

¤ ¤ ¤ ¤ ¤ ¤ ¤

¥ ¥ ¥ ¥ ¥ ¥ ¥

¢ ¢ ¢ ¢ ¢

£ £ £ £ £

¤ ¤ ¤ ¤ ¤ ¤ ¤

¥ ¥ ¥ ¥ ¥ ¥ ¥

¤ ¤ ¤ ¤ ¤ ¤ ¤

¥ ¥ ¥ ¥ ¥ ¥ ¥

¤ ¤ ¤ ¤ ¤ ¤ ¤

¥ ¥ ¥ ¥ ¥ ¥ ¥

¡ ¡ ¡ ¡ ¡

¡ ¡ ¡ ¡ ¡

¡ ¡ ¡ ¡ ¡

¡ ¡ ¡ ¡ ¡

¦ ¦ ¦ ¦ ¦

§ § § § §

¡ ¡ ¡ ¡ ¡

¦ ¦ ¦ ¦ ¦

§ § § § §

¡ ¡ ¡ ¡ ¡

¦ ¦ ¦ ¦ ¦

§ § § § §

¡ ¡ ¡ ¡ ¡

¦ ¦ ¦ ¦ ¦

§ § § § §

¡ ¡ ¡ ¡ ¡

¦ ¦ ¦ ¦ ¦

§ § § § §

¡ ¡ ¡ ¡ ¡

¦ ¦ ¦ ¦ ¦

§ § § §

the figure §

¡ ¡ ¡ ¡ ¡

¡ ¡ ¡ ¡

eight thin necks matched¡

¡ ¡ ¡ ¡

¡ ambient disc B

¡ ¡ ¡ ¡

together ¡

¡ ¡ ¡ ¡ ¡

¡ ¡ ¡ ¡ ¡

¡ ¡ ¡ ¡ ¡

¡ ¡ ¡ ¡ ¡

¡ ¡ ¡ ¡ ¡ Figure 3

Note now that Z 2 area (C) = (a+(t) − a−(t))dt 0 Z 1 = 2 (max Ht − min Ht)dt + 4ε 0 x x = 2l + 4ε.

Thus the image θ(C) can be enclosed by a disc B of area 2l + 10ε (we take 10ε to make up for the extra bits). Step 4. Consider now the symplectic immersion

θ0 = id × θ : R2n × C → R2n × R2.

Obviously θ0(L0) ⊂ R2n+2 is an immersed Lagrangian submanifold and θ0(L0) ⊂ R2n × B. We will show that L00 = θ0(L0) is embedded. Observe that the only place where double points could occur is in the thin necks. But for t ∈ [−ε, ε], gt(L) = L and for t ∈ [1 − ε, 1 + ε], gt(L) = h1(L) and by assumption h1(L)∩L = ∅ so there are no double points and θ0 is an embedding. 34 CHAPTER 3. LAGRANGIAN SUBMANIFOLDS

Step 5. The only thing we are left to show is that L00 is rational so that we can apply 3.2.B. We will prove that γ(L) = γ(L00). Let φ be the composition of the Lagrangian suspension and θ0, that is the map

1 2n ∗ 1 2n 2 φ : L × S → R × T S → R (p1, ..., pn, q1, ..., qn) × R (p, q)

00 sending (x, t) to (gt(x), θ(−G(gt(x), t), t)). Then L is the image of 1 00 L × S under φ. The group H1(L ) is generated by the cycles of the 1 form φ(b) where either b ⊂ L × {0}, or b = {x0} × S for x0 ∈ L. In the first case φ(b) = b×{θ(0, 0)}, so the symplectic areas of b and φ(b) coincide. In the second case write α for the orbit {gtx0}, t ∈ [0; 2]. Note that Z Z Z ∗ ∗ φ (p1dq1 + ... + pndqn + pdq) = p1dq1 + ... + pndqn + θ pdq b α Z = 0 + rdt

Z 2 = − G(gt(x0), t)dt = 0. 0 This proves that L00 is a rational Lagrangian submanifold with γ(L) = γ(L00). Since L00 is contained in R2n × B we get in view of 3.2.B that

γ(L00) ≤ area (B) = 2l + 10ε, 1 for all ε > 0. Therefore e(L) ≥ 2 γ(L). 2 Chapter 4

The ∂¯-equation with Lagrangian boundary conditions

In this chapter we prove Theorem 3.2.B which states that γ(L) ≤ πr2 for any closed rational Lagrangian submanifold L ⊂ B2(r) × R2n−2. The proof is based on Gromov’s techniques of pseudo-holomorphic discs.

4.1 Introducing the ∂¯-operator

2n Identify R (p1, q1, ..., pn, qn) with n n C (p1 + iq1, ..., pn + iqn) = C (w1, ..., wn). Denote by h , i the Euclidean scalar product. The three geometric structures we get in this way are the Euclidean, the symplectic and the complex structure. They are related by the following formula hξ, ηi = ω(ξ, iη). We will check this formula in the case n = 1. Let ξ = (p0, q0) and η = (p00, q00). Then µ ¶ µ ¶ ³ p0 −q00 ´ dp ∧ dq(ξ, iη) = dp ∧ dq , = p0p00 + q0q00 = hξ, ηi. q0 p00

35 36 CHAPTER 4. THE ∂¯-EQUATION

In what follows we will measure areas and lengths using the Euclidean metric. Consider the unit disc D2 ⊂ C with coordinate z = x+iy. For a smooth map f : D2 → Cn we define the ∂¯-operator, ∂¯ : C∞(D2, Cn) → C∞(D2, Cn) by 1³∂f ∂f ´ ∂f¯ = + i . 2 ∂x ∂y

Example. Let f : C → C, z 7→ z.¯ So f(x, y) = x − iy and ∂f¯ = 1 ¯ ∂f 2 (1 + 1) = 1. We observe that ∂f = ∂z¯ . Let us introduce two useful geometric quantities associated with a map f : D2 → Cn. The symplectic area of f is given by Z ω(f) = f ∗ω D2 and the Euclidean area of f is given by Z s D∂f ∂f ED∂f ∂f E D∂f ∂f E2 Area (f) = , , − , dxdy. D2 ∂x ∂x ∂y ∂y ∂x ∂y

Proposition 4.1.A. Z i) Area (f) ≤ 2 |∂f¯ |2dxdy + ω(f), D2 ii) Area (f) ≥ |ω(f)|.

Proof: Given ξ, η ∈ Cn we have the following inequality: p 1 |ξ|2|η|2 − hξ, ηi2 ≤ |ξ||η| ≤ (|ξ|2 + |η|2). 2 But 1 1 1 |ξ + iη|2 + ω(ξ, η) = (|ξ|2 + |η|2) + hξ, iηi + hξ, −iηi = (|ξ|2 + |η|2), 2 2 2 so we get p 1 |ξ|2|η|2 − hξ, ηi2 ≤ |ξ + iη|2 + ω(ξ, η). 2 4.2. THE BOUNDARY VALUE PROBLEM 37

Integrating this pointwise inequality we get 4.1.A.i. To prove 4.1.A.ii it suffices again to verify the pointwise inequality. Assume that η 6= 0 and note that hη, iηi = 0. Projecting ξ on η and iη we get D η E2 D iη E2 ξ, + ξ, ≤ |ξ|2. |η| |iη| Since |η| = |iη| the above reads

hξ, ηi2 + ω(ξ, η)2 ≤ |ξ|2|η|2, hence p |ω(ξ, η)| ≤ |ξ|2|η|2 − hξ, ηi2. 2

4.2 The boundary value problem

Let L ⊂ Cn be a closed Lagrangian submanifold and let g : D2 ×Cn → Cn be a smooth map which is bounded together with all its . n Fix a class α ∈ H2(C ,L). Given this data consider the following problem. Find a smooth map f :(D2, ∂D2) → (Cn,L) such that ½ ∂f¯ (z) = g(z, f(z)) (P (α, g)) [f] = α.

Example. If g = 0, α = 0 then the space of solutions of P (0, 0) consists of the constant mappings f(z) ≡ w for w ∈ L. To see this first observe that ω(f) = 0. Indeed, since α = 0 and L is Lagrangian the curve f(∂D2) bounds a 2-chain in L with zero symplectic area. This chain together with f(D2) forms a closed surface in Cn. Since ω is exact the symplectic area of this surface vanishes. Therefore ω(f) = 0. Further, since g = 0 we get that ∂f¯ = 0. So the first part of 4.1.A ∂f ∂f yields Area (f) = 0 and hence ∂x is parallel to ∂y . On the other hand ∂f ∂f ∂f ∂f ∂f ∂f ∂x = −i ∂y and therefore ∂x ⊥ ∂y . Consequently ∂x = ∂y = 0. So f is 38 CHAPTER 4. THE ∂¯-EQUATION a constant map and because of the boundary condition its image lies in L.

Assume now that we have a sequence of functions {gn} which ∞ C -converges to some function g. Let fn be the solutions of the corre- sponding problems P (α, gn). Gromov’s famous compactness theorem (see [G1], [AL]) states that either {fn} contains a subsequence that converges to a solution of P (α, g) or bubbling off takes place. In order to explain what bubbling off is we introduce the concept of a solution. Definition. Consider the following data:

0 • A decomposition α = α + β1 + ... + βk, where βj 6= 0, j = 1, ...k. • A solution f of P (α0, g).

• Solutions hj of P (βj, 0), these are the so called pseudo-holomorphic discs.

2 2 This object is called a cusp solution of P (α, g) and f(D ) ∪ h1(D ) ∪ 2 ... ∪ hk(D ) is called its image. The bubbling off phenomenon means that there exists a subse- quence of {fn} (which we denote again by {fn}) which converges to a cusp solution of P (α, g). The only feature of this convergence which is important for our purposes is the continuity of the Euclidean area:

Xk Area (fn) → Area (f) + Area (hj). j=1

The complete definition of the convergence is quite sophisticated (see [G1],[AL]), and we omit it. An illustrating example will be given in 4.4 below. Using the compactness theorem, Gromov established the following important result [G1].

Persistence Principle. Consider a “generic” family gs(z, w), s ∈ [0, 1] with g0 = 0. Then either P (0, gs) has a solution for all s or bubbling off occurs at some s∞ ≤ 1, that is there exists a subsequence 4.3. AN APPLICATION TO THE LIOUVILLE CLASS 39

sj → s∞ such that the sequence of solutions of P (0, gsj ) converges to

a cusp solution of P (0, gs∞ ). The word “generic” should be interpreted as follows. One can endow the space of all families gs with an appropriate structure . Generic families form a residual subset (that is a countable intersection of open and dense subsets) in this space. In particular, every family gs becomes generic after an arbitrarily small perturbation. We refer to [G1],[AL] for further details.

4.3 An application to the Liouville class

We present here Sikorav’s proof of Theorem 3.2.B [S1]. Suppose that L ⊂ B2(r) × Cn−1 is a closed Lagrangian submanifold. Take g(z, w) = (σ, 0, ..., 0) ∈ Cn for some σ ∈ C.

Lemma 4.3.A. If |σ| > r then P (0, g) has no solutions.

Proof: Suppose that f is a solution. Denote by φ its first (complex) coordinate. Thus ∂φ ∂φ + i = 2σ. ∂x ∂y ¯ ¯ 2 n−1 ¯ ¯ Since L ⊂ B (r) × C , we have that φ|∂D2 ≤ r. Note that Z ∂φ ∂φ 2πσ = + i dxdy 2 ∂x ∂y ZD = d(φdy − iφdx) ZD2 = φdy − iφdx. S1 Now we can write x+iy = e2πit and dx+idy = 2πie2πitdt, so dy−idx = 2πe2πitdt. Therefore ¯Z 1 ¯ ¯ ¯ 2π|σ| = 2π¯ e2πitφ(e2πit)dt¯ ≤ 2πr 0 and hence |σ| ≤ r. 2 40 CHAPTER 4. THE ∂¯-EQUATION

Take now any σ with |σ| > r and apply the persistence principle to the family gs = (sσ, 0, ..., 0), s ∈ [0, 1]. The previous lemma tells us that there is no solution for s = 1, so we have that for a small perturbation of gs bubbling off takes place. For the sake of simplicity we assume that it happens in gs itself. The general argument goes through without changes (make estimates up to ε) and is left to the reader. So we have a sequence sn → s∞ ≤ 1 and a decomposition 0 =

α + β1 + ... + βk, βj 6= 0. Let fn be the solutions of P (0, gsn ), f∞ a solution of P (α, gs∞ ) and h1, ...hk holomorphic discs with [hj] = βj satisfying Xk Area (fn) → Area (f∞) + Area (hj). j=1

Applying both parts of 4.1.A and using the fact that the discs hj are holomorphic we get that Area (hj) = ω(hj) ≥ γ(L). This inequality follows from the fact that [hj] = βj 6= 0. From 4.1.A.ii we deduce that ¯ ¯ ¯ X ¯ Area (f∞) ≥ |ω(f∞)| = ¯ ω(hj)¯ ≥ γ(L). P Thus Area (f∞) + Area (hj) ≥ 2γ(L). On the other hand 4.1.A.i implies that 2 2 2 Area (fn) ≤ 2πsn|σ| ≤ 2π|σ| . ¯ We use here that ω(fn) = 0 (since [fn] = 0) and ∂fn = gsn . Putting these two inequalities together we get 2π|σ|2 ≥ 2γ(L). This is true for all σ with |σ| > r so we have πr2 ≥ γ(L), which proves the theorem. 2

Proof of 3.2.A: Consider a closed Lagrangian submanifold L ⊂ B2(r) × Cn−1. Then 4.3.A above implies that the problem ½ ∂f¯ (z) = (sσ, 0, ..., 0), |σ| > r [f] = 0. has no solution for s = 1. The persistence principle implies that bub- bling off must take place. This means that there exists a non-zero class β1 which is represented by a holomorphic disc h1. Since h1 6= constant, 4.4. AN EXAMPLE 41

n 2 we get ω(h1) > 0. Since we found a disc in C spanned by h1(∂D ) which has non-zero symplectic area, we conclude that λL 6= 0. 2

4.4 An example

In this section we sketch an elementary example where all the phenom- ena appearing in the proof can be seen explicitly. Let L = ∂D2 ⊂ C and let σ = 1. We wish to find all maps f : D2 → C such that f(∂D2) ⊂ ∂D2 and ½ ∂f¯ (z, z¯) = s (4.4.A) [f|∂D2 ] = 0. ∂f Since ∂z¯ = s then f(z, z¯) = sz¯+u(z) where u is a holomorphic function on D2. We claim that the function s+zu(z) is holomorphic, takes ∂D2 2 to ∂D and that (s + zu(z))|∂D2 has degree 1. Indeed,

zf(z, z¯) = s|z|2 + zu(z) so ¯ ¯ |z| |f(z, z¯)| = ¯s|z|2 + zu(z)¯. If |z| = 1 this reduces to |f(z, z¯)| = |s + zu(z)| but |f(z, z¯)| = 1. So s + zu(z) is a holomorphic function taking ∂D2 to ∂D2. Observe that deg f = 0 and deg z = 1 so the total degree, deg zf = 1 and hence deg (s + zu(z)) = 1. All such holomorphic functions are known as isometries of the hyperbolic metric in the unit disc. They have the form 1 − αz¯ eiθ z − α iθ 1−αz¯ for θ ∈ R and |α| > 1. Thus s + zu(z) = e z−α so

eiθ + αs − z(s + eiθα¯) zu(z) = . z − α Since u is holomorphic it cannot have any poles, so eiθ + αs = 0 and eiθ 1 hence α = − s . Now 1 < |α| = | s | implies that s < 1 and that there 42 CHAPTER 4. THE ∂¯-EQUATION

are no solutions of 4.4.A for s ≥ 1. So bubbling off must occur at s = 1. For the sake of simplicity we put θ = 0. Then

1 2 s − s 1 − s u(z) = − 1 = z + s sz + 1

1−s2 so fs(z, z¯) = sz¯ + sz+1 . When s goes to 1, fs(z, z¯) → z¯ for all z 6= −1 and this convergence is uniform outside every neighbourhood of −1. 2 2 Consider the graphs of fs in D × D ⊂ C × C. Set w = fs(z, z¯) so (w − sz¯)(sz + 1) = 1 − s2. When s → 1 this equation goes to (w − z¯)(z + 1) = 0 and the graph becomes the of two curves ½ w =z ¯ z = −1.

Here w =z ¯ is the graph of f∞ and {z = −1} corresponds to the holomorphic disc with boundary on {−1} × L. Projecting onto the w-coordinate we get bubbling off. Indeed, f∞(z) =z ¯ is a solution of P (−a, 1) where a = [S1] and the holomorphic disc h(z) = z is a solution of P (a, 0). In order to visualize the bubbling off phenomenon we restrict to the real axes. Consider the graphs of the corresponding functions 1−s2 fs(x) = sx + sx+1 for x ∈ [−1, 1]. We get the following picture, see figure 4. The graphs of fs converge to the union of two curves, the graph of the real part of f∞ and the segment I = [−1, 1] which is the real part of the holomorphic disc {−1} × D2. 4.4. AN EXAMPLE 43

w graph(fs ) 1 image(h) graph(f )

-1 1 x

-1

Figure 4 44 CHAPTER 4. THE ∂¯-EQUATION Chapter 5

Linearization of Hofer’s geometry

In this chapter we interpret the function ρ(1l, φ) as the distance be- tween a point and a subset in a linear normed space. Later on this will enable us in some interesting cases to find lower bounds for ρ(1l, φ).

5.1 The space of periodic Hamiltonians

Denote by F the space of all smooth normalized Hamiltonian functions F : M × R → R which are 1-periodic in time: F (x, t + 1) = F (x, t) for all x ∈ M. We will often consider such F as functions on M ×S1, where 1 S = R/Z. For a function F ∈ F denote by φF the time-one-map f1 of the corresponding Hamiltonian flow {ft}. Note that every Hamiltonian diffeomorphism φ can be represented in this way. Indeed, let {gt} be any flow with g1 = φ. Take a function a : [0; 1] → [0; 1] such that a ≡ 0 near 0 and a ≡ 1 near 1. Consider the new flow ft = ga(t) and extend it on the whole R by the formula ft+1 = ftf1. Clearly we get a smooth flow. The claim follows from the next exercise.

Exercise 5.1.A. Prove that a flow {ft}, t ∈ R is generated by a function from F if and only if ft+1 = ftf1 for all t.

45 46 CHAPTER 5. LINEARIZATION

Consider a subset H ⊂ F defined by

H = {H ∈ F | φH = 1l}.

In other words Hamiltonians from H generate loops of Hamiltonian diffeomorphisms (or Hamiltonian loops). Introduce a norm on F as follows:

|||F ||| = max ||Ft|| = max(max F (x, t) − min F (x, t)). t t x x Now we are ready to present the main theorem of this chapter. Theorem 5.1.B. For every F ∈ F

ρ(1l, φF ) = inf |||F − H|||. H∈H

F H . F. 0

ρ(1l, φF ) Figure 5

Note that the right hand side is simply the distance from F to H in the sense of our norm (see figure 5). Thus the set H carries a lot of information about Hofer’s geometry. In the next chapters we will establish some interesting properties of H and will look closely at Hamiltonian loops. Theorem 5.1.B is a simple consequence of the following fact. Lemma 5.1.C. For every φ ∈ Ham(M)

ρ(1l, φ) = inf |||F |||, 5.1. THE SPACE OF PERIODIC HAMILTONIANS 47

where the infimum is taken over all Hamiltonians from F which gen- erate φ.

In the terminology of some of my papers this means that the ”coarse” Hofer’s norm coincides with the usual one. Proof of 5.1.C: For φ ∈ Ham(M, Ω) set r(1l, φ) = inf |||F ||| where F runs over all Hamiltonians F ∈ F which generate φ. Clearly r(1l, φ) ≥ ρ(1l, φ). Our task is to prove the converse inequality. Fix a positive number ². Choose a path {ft}, t ∈ [0; 1] of Hamiltonian diffeomor- R 1 phisms such that f0 = 1l, f1 = φ and such that 0 m(t)dt ≤ ρ(1l, φ) + ² where m(t) = ||Ft||. We can assume without loss of generality that F ∈ F and that m(t) > 0 for all t. Indeed, to guarantee the periodicity one makes a time reparametrization as it is explained in the beginning of this section. The justification of the second assumption is given in the next section. Denote by C the space of all C1-smooth orientation preserving diffeomorphisms of S1 which fix 0. Note that for a ∈ C the path fa = {fa(t)} is generated by the normalized Hamiltonian function F a(x, t) = a0(t)F (x, a(t)), where a0 denotes the derivative with respect to t (see 1.4.A above). Take now a(t) as the inverse of

R t 0 m(s)ds b(t) = R 1 . 0 m(s)ds Note that

Z 1 |||F a||| = max a0(t)m(a(t)) = max(m(t)/b0(t)) = m(t)dt. t t 0 We conclude that |||F a||| ≤ ρ(1l, φ) + ². Approximating a in the C1- topology by a smooth diffeomorphism from C we see that one can find a smooth normalized Hamiltonian, say F˜, which generates φ and such that |||F˜||| ≤ ρ(1l, φ) + 2². Since this can be done for an arbitrary ² we conclude that r(1l, φ) ≤ ρ(1l, φ). This completes the proof. 2

Proof of 5.1.B: Write {ft} for the Hamiltonian flow generated by F . Let {gt} be any other Hamiltonian flow generated by G ∈ F with g1 = φF . Decompose gt as ht ◦ ft. It follows from 5.1.A that {ht} is a loop 48 CHAPTER 5. LINEARIZATION

of Hamiltonian diffeomorphisms, that is its normalized Hamiltonian H belongs to F and h0 = h1 = 1l. Vice versa, for every loop {ht} the 1 flow {ht ◦ ft} is generated by a Hamiltonian from F, and its time one map equals φF . Further,

−1 G(x, t) = H(x, t) + F (ht x, t).

0 0 −1 Set H (x, t) = −H(htx, t). Note that H generates a loop {ht }, and thus belongs to H. On the other hand the expression for G above implies that |||G||| = |||F − H0|||. In view of the discussion above to each G corresponds unique H0 and vice versa. Thus the required statement follows immediately from Lemma 5.1.C. 2

5.2 Regularization

Here we justify the assumption m(t) > 0 made in the proof of Lemma 5.1.C. A flow {ft} is called regular if for every t the normalized Hamil- tonian function Ft does not vanish identically. In other words, for every t the tangent vector to the path {ft} does not vanish.

Proposition 5.2.A. Let {ft} be a flow generated by a Hamiltonian from F . Then there exists an arbitrary small (in the C∞-sense) loop −1 {ht} such that the flow {ht ft} is regular. The proof is divided in several steps.

1) Let us analyze the problem. Assume that {ht} is a loop gen- erated by a Hamiltonian H ∈ H. Then the Hamiltonian of the flow −1 {ht ft} is given by −H(htx, t) + F (htx, t). We have to prove that for every t this expression does not vanish identically. But this is equiva- lent to the assertion that

F (x, t) − H(x, t) 6≡ 0 (5.2.B) for all t. So our task is to produce an arbitrary small H ∈ H which satisfies 5.2.B.

1 Note that in general the flow {ft ◦ ht} (mind the order) is not generated by a periodic Hamiltonian! 5.2. REGULARIZATION 49

2) Let us introduce another useful notion. A k-parameter variation of the constant loop is a smooth family of loops {ht(²)}, where ² belongs k to a neighbourhood of 0 in R and ht(0) = 1l for all t. When M is open we assume in addition that the supports of all ht(²) are contained in some compact subset of M. Here is a convenient way to produce variations. Let us start with the 1-parameter case. Take a function G ∈ F such that

Z 1 G(x, t)dt = 0 (5.2.C) 0 for every x ∈ M. Then define ht(²) ∈ Ham(M, Ω) as the time-²-map of the Hamiltonian flow generated by the time-independent Hamiltonian R t function 0 G(x, s)ds. Exercise 5.2.D. Let H(x, t, ²) be the normalized Hamiltonian func- tion of the loop {ht(²)}. Show that ∂ | H(x, t, ²) = G(x, t). ∂² ²=0

It is natural to construct k-parameter variations as compositions of 1-parameter variations:

(1) (k) ht(²1, ..., ²k) = ht (²1) ◦ ... ◦ ht (²k). Here every h(j) is constructed with the help of a function G(j) as above. Exercise 5.2.D yields that the of the Hamil- (j) tonian H(x, t, ²) with respect to ²j at ² = 0 equals G . 3) Fix a point y ∈ M, and consider the 2n-dimensional linear ∗ space E = Ty M. Choose 2n smooth closed curves α1(t), ..., α2n(t) (where t ∈ S1) which satisfy the following conditions:

R 1 • 0 αj(t)dt = 0 for all j = 1, ..., 2n;

• the vectors α1(t), ..., α2n(t) are linearly independent for every t. Here is a construction of such a system of curves. Choose a basis u1, v1, ..., un, vn in E and take curves of the form uj cos 2πt + vj sin 2πt and −uj sin 2πt + vj cos 2πt. 50 CHAPTER 5. LINEARIZATION

4) Choose now functions G1(x, t), ..., G2n(x, t) from F which sat- (j) isfy condition 5.2.C above and such that dyGt = αj(t). Take the corresponding 2n-parameter variation {ht(²)} of the constant loop as 1 2n in step 2. Consider the mapping Φ : S × R (²1, ..., ²2n) → E defined by (t, ²) → dy(Ft − Ht(²)). It follows that Φ is a in some neighbourhood U of the circle {² = 0}. Indeed, our construction together with the discussion in step 2 imply that ∂ |²=0Φ(t, ²) = αj(t). ∂²j But these vectors generate the whole E. Denote by Ψ the restriction of Φ to S1 × U. Since Ψ is a submersion, the set Ψ−1(0) is a one- dimensional submanifold of S1 × U, so its projection to U is nowhere dense. Thus there exist arbitrary small values of the parameter ² such that dy(Ft − Ht(²)) 6= 0 for all t. Therefore for every t the condition 5.2.B above is satisfied. This completes the argument. 2

5.3 Paths in a given homotopy class

A homotopy is simply a smooth 1-parameter family of paths. We will consider homotopies of non-closed paths with fixed end points, as well as homotopies of loops based at 1l, unless otherwise stated. When the manifold M is open we as usually assume that supports of all diffeomorphisms appearing in 2-parameter families are contained in an ambient compact subset of M. Take a function F ∈ F, and denote by {ft} the corresponding Hamiltonian flow. Consider the quantity

l(F ) = inf length{gt},

where the infimum is taken over all Hamiltonian paths {gt}, t ∈ [0; 1] with g0 = 1l, g1 = φF which are homotopic to {ft} with fixed end points. Let us sketch a useful interpretation of this quantity. Consider 5.3. PATHS IN A GIVEN HOMOTOPY CLASS 51 the universal cover Z of (Ham(M, Ω), 1l). It is defined in the standard manner with the minor exception that we consider smooth paths and smooth homotopies only. The Finsler structure on Ham(M, Ω) lifts canonically to the universal cover. This gives rise to the notion of the length of a smooth curve on Z, and therefore to the distance function ˜ ˜ ρ˜ on Z. Denote by 1l the canonical lift of 1l to Z, and by φF the lift of φF associated to the path {ft}, t ∈ [0; 1]. With this language ˜ ˜ l(F ) =ρ ˜(1l, φF ), so this quantity is responsible for the geometry of the universal cover. Denote by Hc the set of all Hamiltonians from H which generate contractible loops. In other words Hc is the path connected component of 0 in H. Theorem 5.3.A. For every F ∈ F

l(F ) = inf |||F − H|||. H∈Hc This can be proved exactly in the same way as Theorem 5.1.B above. In the course of the proof one should take into account the following simple additional observations:

• Time reparametrizations as well as the regularization procedure 5.2 do not change the homotopy class of a path with fixed end points. Thus one reduces the problem to the minimization of |||G||| over all G ∈ F with φG = φF whose Hamiltonian flow is homotopic to {ft} (cf. 5.1.C above).

• If gt = ht ◦ ft, where {ft} and {gt} are homotopic with fixed end points, then the loop ht is contractible (cf. the final argument in the proof of 5.1.B).

The details of the proof are left to the reader. 52 CHAPTER 5. LINEARIZATION Chapter 6

Lagrangian intersections

The theory of Lagrangian intersections studies one of the most surpris- ing phenomena in symplectic topology. In this chapter we review some results from this theory which being combined with the linearization idea above give a rather powerful tool for investigation of the geometry of the group of Hamiltonian diffeomorphisms.

6.1 Exact Lagrangian isotopies

Let (V 2n, ω) be a symplectic manifold and let N n be a closed manifold. Let Φ: N × [0, 1] → V be a smooth family of Lagrangian embeddings i.e. Φ is a Lagrangian ∗ isotopy. Note that Φ ω must be of the form αs ∧ ds where {αs} is a family of 1-forms on N (since Φ∗ω vanishes on the fibers N ×{point} ). ∗ Furthermore, observe that dΦ ω = dαs ∧ ds = 0 which implies that αs is closed for all s.

Definition. A Lagrangian isotopy Φ is exact if αs is exact for all s. Exercise 6.1.A. Show that a Lagrangian isotopy is exact if and only if it can be extended to an ambient Hamiltonian isotopy of V . Hint: −1 Write αs = dHs on N and extend Hs ◦ Φs to a time dependent normalized Hamiltonian function on V .

53 54 CHAPTER 6. LAGRANGIAN INTERSECTIONS

Example. Let V be a surface and N = S1. Then Φ is exact if and only if the oriented area between Φ(N ×{0}) and Φ(N ×{s}) vanishes for all s. See figure 6 for the case V = S2, and figure 7 for the case V = T ∗S1 = R×S1. Note that in the case of the cylinder it is possible to find a symplectic isotopy describing the right hand side picture.

exact not exact Figure 6

exact not exact Figure 7

The next result plays an important role in our study of Hofer’s geometry. Assume that {ht} is a loop of Hamiltonian diffeomorphisms of (M, Ω) generated by H ∈ H. Let L ⊂ M be a closed Lagrangian submanifold. Consider the Lagrangian suspension (see 3.1.E above).

L × S1 → (M × T ∗S1, Ω + dr ∧ dt), 6.1. EXACT LAGRANGIAN ISOTOPIES 55

(x, t) 7→ (htx, −H(htx, t), t). Our aim is to investigate the behaviour of this Lagrangian embedding under a 1-parameter deformation. Let {ht,s}, s ∈ [0; 1] be a smooth family of Hamiltonian loops. Denote by

Φ: L × S1 × [0; 1] → M × T ∗S1

the corresponding family of Lagrangian suspensions. Theorem 6.1.B. The Lagrangian isotopy Φ is exact. In other words, a homotopy of Hamiltonian loops gives rise to an exact Lagrangian isotopy of Lagrangian suspensions. The proof of the theorem is based on the following auxiliary result. Denote by H(x, t, s) the normalized Hamiltonian of the loop {ht,s}. Proposition 6.1.C. For every x ∈ M and s ∈ [0; 1] Z 1 ∂H (ht,sx, t, s)dt = 0. 0 ∂s Proof of the Proposition: The proof is based on the following for- mula, which is valid for an arbitrary 2-parameter family of diffeomor- phisms on a manifold. We leave its verification to the reader (see also [B1]). Consider vector fields Xt,s and Yt,s on M such that d h x = X (h x), dt t,s t,s t,s and d h x = Y (h x). ds t,s t,s t,s Then ∂ ∂ X = Y + [X ,Y ]. ∂s t,s ∂t t,s t,s t,s

Note that for every t and s fields Xt,s and Yt,s are Hamiltonian vector fields. Of course Xt,s = sgradHt,s where H is already introduced above. Write Yt,s = sgradFt,s. Recall that

[sgradH, sgradF ] = −sgrad{H,F }. 56 CHAPTER 6. LAGRANGIAN INTERSECTIONS

Thus we get that ∂H ∂F ∂F t,s = t,s − {H ,F } = t,s + dF (sgradH ). ∂s ∂t t,s t,s ∂t t,s t,s

But the last expression evaluated at point ht,sx equals d F (h x, t). dt t,s We conclude that ∂H t,s (h x) ∂s t,s is the full derivative of a , thus its integral over the period must vanish. This completes the proof. 2

∗ Proof of 6.1.B: Write Φ (Ω + dr ∧ dt) as αs ∧ ds. We have to check that αs is an exact form. The form αs can be calculated explicitly. Exercise (cf. 3.1.E). Show that ∂h α (ξ) = Ω(h ξ, t,s x) s t,s∗ ∂s for all x ∈ L, ξ ∈ TxL, and ∂ ∂H α ( ) = (h x, t, s). s ∂t ∂s t,s

1 Note that the first homology group H1(L × S , Z) is generated by split cycles of the form C = β × {0} and D = {y} × S1. Here β is a cycle on L, and y is a point of L. In order to prove that the form αs is exact it suffices to verify that its integrals over all 1-cycles vanish. For the cycles of the form C this follows from the fact that h0,s ≡ 1l for all s. Thus the exercise above implies that αs vanishes on all vectors tangent to L × {0}. Further, Z Z 1 ∂H αs = (ht,sy, t, s)dt. D 0 ∂s This expression vanishes in view of Proposition 6.1.C above. This com- pletes the proof. 2 6.2. LAGRANGIAN INTERSECTIONS 57 6.2 Lagrangian intersections

We say that a Lagrangian submanifold N ⊂ V has the Lagrangian intersection property if N intersects its image under any exact La- grangian isotopy. In view of 6.1.A above one can reformulate this as N ∩ φ(N) 6= ∅ for all φ ∈ Ham(V, ω), or in other words the displace- ment energy of N is infinite: e(N) = +∞. Examples 6.2.A. Infinitesimal Lagrangian intersection problem Let F be an autonomous Hamiltonian function on V and let ξ = sgradF be its Hamiltonian vector field. Then ξ is tangent to N exactly at critical points of F |N (exercise). Since N is closed, F |N must have critical points and hence one cannot displace N by an infinitesimal Hamiltonian isotopy. 6.2.B. Gromov’s Theorem If π2(V,N) = 0 and V has “nice” behaviour at infinity (say V is a product of a closed manifold and a cotangent bundle), Gromov [G1] (see also Floer [F]) showed that N has the Lagrangian intersection property. In particular, this applies to the circle {r = 0} in T ∗S1 (of course, this can be shown by an elementary area control, see figure 7 and the discussion above). More generally, this holds for every non- contractible curve on an oriented surface. A proof of Gromov’s theorem was sketched in 3.2.G above. We refer to [AL, Chap. X] for the details. If π2(V,N) 6= 0, the Lagrangian intersection property can be vio- lated. Take for example a circle N of tiny area on V = S2 and displace it, see figure 8.

However, the Lagrangian intersection property obviously holds for the equator (use that the equator divides the sphere into two discs of equal area). Definition 6.2.C. Let L be a closed Lagrangian submanifold of a symplectic manifold (M, Ω). We say that L has the stable Lagrangian intersection property if L × {r = 0} has the Lagrangian intersection property in (M × T ∗S1, Ω + dr ∧ dt). 58 CHAPTER 6. LAGRANGIAN INTERSECTIONS

Figure 8

Let us give two examples which play an important role in the next chapters. 6.2.D. Tori in T ∗Tn. Consider a Lagrangian torus in the cotangent bundle T ∗Tn endowed with the standard symplectic structure (see 3.1.C above). Assume that it is homologous to the zero section. It is easy to check that the topological assumption of Theorem 6.2.B holds in this situation. Therefore such tori have the stable Lagrangian intersection property. 6.2.E The equator of S2. The stable Lagrangian intersection property also holds for equators of S2. This follows from a difficult theorem due to Oh ([O1], [O2]) based on a clever version of Floer homology. I do not know an example of a closed connected Lagrangian submani- fold which has the Lagrangian intersection property, but does not have the stable one.

6.3 An application to Hamiltonian loops

Let (M, Ω) be a symplectic manifold. Assume that L ⊂ M is a closed Lagrangian submanifold which has the Lagrangian intersection prop- erty. Let {gt} be a loop of Hamiltonian diffeomorphisms generated by a Hamiltonian G ∈ H. Assume in addition that

1 • gt(L) = L for all t ∈ S ; • G(x, t) = 0 for all x ∈ L, t ∈ S1. 6.3. AN APPLICATION TO HAMILTONIAN LOOPS 59

An obvious example of such a loop is the constant loop gt ≡ 1l. A less straightforward example will be given in 6.3.C below.

Theorem 6.3.A. Let {ht} be any other loop of Hamiltonian diffeo- morphisms homotopic to {gt} as above. Let H ∈ H be its Hamiltonian function. Then there exist x ∈ L and t ∈ S1 such that H(x, t) = 0. As we will see in the next chapter, this result gives rise to non- trivial lower bounds for Hofer’s distance. Proof: Apply twice the Lagrangian suspension construction to L, first using {gt} and then using {ht}. Denote by NG and NH the correspond- ing Lagrangian submanifolds of M × T ∗S1. From the formula for the Lagrangian suspension we learn that NG = L × {r = 0}. Theorem 6.1.B above implies that NH is isotopic to NG by an exact Lagrangian isotopy. Thus the stable Lagrangian intersection property yields that NH ∩ NG =6 ∅. Let (x, 0, t), x ∈ L be the intersection point. Since it lies on NH we have that x = hty and 0 = −H(hty, t, 0) for some y ∈ L. We conclude that H(x, t) = 0. 2

Recall that Hc denotes the space of all 1-periodic Hamiltonians which generate a contractible loop of Hamiltonian diffeomorphisms. As an immediate consequence of the theorem above we get the following result. Corollary 6.3.B. Let L ⊂ M be a closed Lagrangian submanifold which has the stable Lagrangian intersection property. Then for every 1 H ∈ Hc there exist x ∈ L and t ∈ S such that H(x, t) = 0.

For instance this is true when M = S2 and L is the equator of S2. Note that the statement of the corollary in general becomes wrong if we do not assume the (stable) Lagrangian intersection property. 3 Example 6.3.C. Consider the Euclidean space R (x1, x2, x3). Let M = S2 be the unit sphere in the space endowed with the induced area form. The full 1-turn rotation around the x3-axis is a Hamiltonian loop generated by the normalized Hamiltonian function F1(x) = 2πx3. (The reader can check this using our calculation in 1.4.H above.) Thus the k-turn rotation is generated by Hamiltonian Fk(x) = 2πkx3. Since Fk vanishes identically on the equator L = {x3 = 0}, it must vanish 60 CHAPTER 6. LAGRANGIAN INTERSECTIONS

on every simple closed curve on S2 which divides the sphere into two parts of equal areas. Note now that when k is even the k-turn rota- tion of S2 represents a contractible loop in SO(3), and therefore in 2 Ham(S ). On the other hand Fk ≡ 2πk² on the circle C² = {x3 = ²}. We conclude that the phenomenon described in 6.3.B above is very rigid. It disappears completely when one considers circles which divide the sphere into parts of arbitrarily close but non-equal areas. Indeed, given any positive ², choose k such that Fk restricted to C² is arbitrar- ily large! The crucial point, of course, is that the circle C² does not have the Lagrangian intersection property. One can displace it by an element of SO(3) (cf. figure 8 above). Chapter 7

Diameter

In the present chapter we prove that the group of Hamiltonian dif- feomorphisms of a closed oriented surface has infinite diameter with respect to Hofer’s metric.

7.1 The starting estimate

Let (M, Ω) be a symplectic manifold, and let L ⊂ M be a closed Lagrangian submanifold with the stable Lagrangian intersection prop- erty. Let F ⊂ F be a Hamiltonian function such that |F (x, t)| ≥ C for all x ∈ L and t ∈ S1. Here C is a positive constant. The next proposi- ˜ ˜ tion gives a lower bound for the quantity l(F ) =ρ ˜(1l, φF ) introduced in 5.3 above. Proposition 7.1.A. Under these assumptions l(F ) ≥ C.

Proof: This is an immediate consequence of 5.3.A and 6.3.B above. Indeed, 6.3.B states that every function H ∈ Hc vanishes at some point (y, τ) where y ∈ L and τ ∈ S1. Thus |F (y, τ) − H(y, τ)| ≥ C and hence |||F − H||| ≥ C. Since this holds true for every H ∈ Hc we conclude from 5.3.A that l(F ) ≥ C. 2

We wish to extend the estimate above to ρ(1l, φF ). Note that if the group Ham(M, Ω) is simply connected with respect to the (strong Whitney) C∞-topology then all paths with common end points are

61 62 CHAPTER 7. DIAMETER

homotopic and therefore l(F ) = ρ(1l, φF ). If the fundamental group π1(Ham(M, Ω)) is non-trivial then in general there is no hope to extend estimate 7.1.A without some additional information. Indeed, it could happen that in the situation described above there exists a shorter path joining 1l with φF , which of course cannot be homotopic to the flow {ft}, t ∈ [0; 1] generated by F . It turns out however that in some interesting cases one can go round this difficulty. In order to do this we have to look more closely at the fundamental group of Ham(M, Ω).

7.2 The fundamental group

Not much is known about π1(Ham(M, Ω)). There is a complete picture for surfaces (based on classical methods) and some four-dimensional manifolds [G1],[A],[AM] (based on Gromov’s theory of pseudo-holomor- phic curves). In higher dimensions, only a few partial results are avail- able (actually, I do not know any symplectic manifold M of dimension ≥ 6 for which one can completely describe the fundamental group of Ham(M, Ω)). For instance, it is known that Ham(R2n) is simply con- nected (and in fact, contractible) for n = 1, 2, while nothing is known already for n = 3. We will need the information about π1(Ham(M, Ω)), where M is a closed orientable surface. 7.2.A. The sphere (cf. 1.4.H and 6.3.C above). The inclusion SO(3) → Ham(S2) induces the of fundamental groups. In partic- 2 ular, π1(Ham(S )) equals Z2. The non-trivial element is generated by the 1-turn rotation of the sphere around the vertical axis. 7.2.B. Surfaces of genus ≥ 1. In this case one can show that the group of Hamiltonian diffeomorphisms is simply connected. These facts are well known to experts, but it seems to me that they have not appeared in published form. Thus I will outline the argument and give some auxiliary references with the hope that this will enable the reader to reconstruct the proof. We write Diff0(M) (resp. Symp0(M)) for the of the group of all (resp. symplectic) diffeomorphisms of surface M. Sketch of the proof: 7.2. THE FUNDAMENTAL GROUP 63

1)The inclusion π1(Symp0(M)) → π1(Diff0(M)) is an isomor- phism. To see this, consider the space X of all area forms on M with total area 1. Fix an area form Ω ∈ X. Consider the mapping ∗ Diff0(M) → X which sends a diffeomorphism f to the form f Ω. One can adjust the proof of Moser’s stability theorem [MS, pp.94-95] in order to show that this mapping is a Serre fibration. Note that its

fiber is nothing else but Symp0(M, Ω), and the base X is contractible. The desired fact now follows from the homotopy exact sequence of the fibration.

2)The topology of Diff0(M) is known (see [EE]). In particular, this group is contractible for surfaces of genus ≥ 2. Further, when M = S2 it has SO(3) as strong deformation retract, and when M = T2 it has T2 as strong deformation retract (here the torus acts on itself by translations).

3) The inclusion j : π1(Ham(M, Ω)) → π1(Symp0(M, Ω)) is injec- tive (see [MS], 10.18 (iii)). In fact this is true for all closed symplectic manifolds. 4) Let us combine these facts. We immediately get 7.2.B for sur- 2 2 faces of genus ≥ 2. Taking into account that Symp0(S ) = Ham(S ) (see 1.4.C) yields 7.2.A. It remains to handle the case of the torus T2. 2 2 5) Fix a point y ∈ T , and consider the evaluation map Diff0(T ) → 2 T which sends a diffeomorphism f to f(y). It induces a map eD : 2 2 π1(Diff0(T )) → π1(T ). It is easy to see from step 2 that eD is an isomorphism. Consider now the restrictions of the evaluation map to 2 2 Ham(T ) and Symp0(T ), and denote by eH and eS respectively the induced of the fundamental groups. Step 1 implies that eS is an isomorphism. Using step 3 we get that eH = eS ◦j, where j is injective. It follows from a theorem by Floer that eH vanishes (see 2 [LMP1]). Thus π1(Ham(T )) = 0. The proof is complete. 2

Theorem 7.2.C. ([P5]) Assume that

• either M = S2 and L ⊂ S2 is an equator,

• or M is a closed orientable surface of genus ≥ 1 and L is a non-contractible closed curve. 64 CHAPTER 7. DIAMETER

Let F ∈ F be a function such that |F (x, t)| ≥ C for all x ∈ L and 1 t ∈ S , where C is a positive constant. Then ρ(1l, φF ) ≥ C.

Proof: If M has genus ≥ 1 then l(F ) = ρ(1l, φF ) since the group Ham(M, Ω) is simply connected (use 7.2.B). Thus the result follows 2 from 7.1.A above. If M = S then the non-trivial element of π1(Ham(M, Ω)) is represented by the 1-turn rotation (see 7.2.A). Its Hamiltonian van- ishes on L (see 6.3.C). Thus Theorem 6.3.A implies that every function from H must vanish at some point of L × S1. The needed statement now follows from 5.1.B above. 2

Corollary 7.2.D. The group of Hamiltonian diffeomorphisms of a closed surface has infinite diameter with respect to Hofer’s metric.

Indeed, let M and L be as in 7.2.C above, and let B ⊂ M be an open disc disjoint from L. Take a time-independent Hamiltonian F ∈ F which is identically equal to C outside B. The theorem above implies that ρ(1l, φF ) ≥ C. Taking C arbitrarily large we get that the diam- eter is infinite. Note also that in this example the support of φF is contained in B. Thus shrinking B and simultaneously increasing C we get a sequence of Hamiltonian diffeomorphisms which converges to the identity in the C0-sense but diverges in Hofer’s metric. Let me mention that for a closed surface M of genus ≥ 1 there exist at least two other different proofs of the fact that the diameter of Ham(M, Ω) is infinite. One of them goes as follows. Exercise 7.2.E. (see [LM2]). Let F be a normalized Hamiltonian on M. Assume that some regular level set of F contains a non-contractible ˜ closed curve. Consider the lift ft of the corresponding Hamiltonian ˜ flow ft to the universal cover M of M. Show that there exists ² > 0 ˜ ˜ and a family of discs Dt ⊂ M of area ²t such that ftDt ∩ Dt = ∅ for all sufficiently large t. Conclude from Theorem 3.2.C above that ρ(1l, ft) → +∞ when t → +∞. Another proof (see [Sch3]) is based on Floer homology (see Chapter 13 for applications of Floer homology to Hofer’s geometry). 7.3. THE LENGTH SPECTRUM 65

Finally, the fact that diam Ham(M, Ω) = +∞ has been proved for some higher-dimensional manifolds (see [LM2], [Sch3], [P5]). In general, however, this is still an open problem.

7.3 The length spectrum

As we have seen above, our method provides a lower bound for ρ(1l, φF ) when we have very precise information on the fundamental group of Ham(M, Ω). It was already mentioned that no such information is available in dimension ≥ 6. Now we will slightly modify the way of arguing and extend the class of manifolds where the method works. The next notion is one of the main characters of our story.

Definition 7.3.A. For an element γ ∈ π1(Ham(M, Ω)) define the norm of γ by ν(γ) = inf length{ht}, where the infimum is taken over all Hamiltonian loops {ht} which represent γ. The set

{ν(γ)|γ ∈ π1(Ham(M, Ω))} is called the length spectrum of Ham(M, Ω). Exercise. (i) Show that π1(Ham(M, Ω)) is always an abelian group (use the same argument as for finite-dimensional Lie groups). Thus we denote the group operation by +, and write 0 for the neutral element. (ii) Prove that ν(γ) = ν(−γ), and ν(γ + γ0) ≤ ν(γ) + ν(γ0). No general statement about the non-degeneracy of ν is known at the moment (and I would not be surprised by an example of γ 6= 0 with ν(γ) = 0, cf. example 7.3.B below). Thus, frankly speaking, ν is a pseudo-norm, though we call it norm. In Chapter 9 below we will describe a method which provides lower bounds for ν(γ), and in particular enables us to calculate the length spectrum for S2. Example 7.3.B. Liouville manifolds. We say that an open sym- plectic manifold (M, Ω) has the Liouville property if there exists a 66 CHAPTER 7. DIAMETER

smooth family of diffeomorphisms

Dc : M → M, c ∈ (0; +∞)

∗ such that D1 = 1l and Dc Ω = cΩ for all c. Such diffeomorphisms Dc are of course not compactly supported for c 6= 1. An important class of examples is given by cotangent bundles endowed with the standard symplectic structure (see 3.1.C above). The diffeomorphism Dc in this situation is just the fiberwise homothety (p, q) → (cp, q). We claim that the length spectrum of Ham(M, Ω) equals {0} provided (M, Ω) has the Liouville property. The proof of the claim is based on the following simple fact.

Exercise. Let {ft} be the Hamiltonian flow generated by a normalized −1 Hamiltonian F (x, t). Then for every c > 0 the flow {DcftDc } is again −1 a Hamiltonian flow whose normalized Hamiltonian equals cF (Dc x, t).

Let {ht} be a loop of Hamiltonian diffeomorphisms. Consider the fam- −1 ily of homotopic loops {DchtDc }. It follows from the exercise above that the lengths of the loops go to zero when c → 0, thus every loop can be homotoped to a loop of an arbitrary small length. We conclude that the length spectrum is {0} (without any knowledge about the fundamental group ). In the next chapter we will discuss this example in the context of classical mechanics.

7.4 Refining the estimate

Theorem 7.4.A. Let (M, Ω) be a symplectic manifold and let L ⊂ M be a closed Lagrangian submanifold with the stable Lagrangian intersection property. Assume that the length spectrum of Ham(M, Ω) is bounded from above by some K ≥ 0. Let F ∈ F be a function such that |F (x, t)| ≥ C for all x ∈ L and t ∈ S1. Then

ρ(1l, φF ) ≥ C − K.

Proof: Fix an arbitrary ² > 0. Let {gt} be a path of Hamiltonian −1 diffeomorphisms which joins 1l with φF . Consider the loop {ft ◦ gt }. 7.4. REFINING THE ESTIMATE 67

By our assumption this loop is homotopic to a loop {ht} whose length does not exceed K + ². The path {ft} is homotopic with fixed end points to the composition {ht ◦ gt}. Therefore

l(F ) ≤ length{ht} + length{gt}.

Since l(F ) ≥ C in view of 7.1.A we conclude that length{gt} ≥ C − K − ². Thus ρ(1l, φF ) ≥ C − K. 2 68 CHAPTER 7. DIAMETER Chapter 8

Growth and dynamics

In this chapter we discuss the asymptotic geometric behaviour of one- parameter subgroups of the group of Hamiltonian diffeomorphisms. We describe a link between geometry and invariant tori of classical mechanics.

8.1 Invariant tori of classical mechanics

Invariant Lagrangian tori of Hamiltonian dynamical systems play an important role in classical mechanics. We start with an obstruction to the existence of invariant tori which comes from the geometry of the group of Hamiltonian diffeomorphisms (see 8.1.C below). Consider the n-dimensional torus Tn endowed with the Euclidean 2 n 2 metric ds = Σj=1dqj . The Euclidean geodesic flow can be described by a Hamiltonian system on the cotangent bundle T ∗Tn endowed with the standard symplectic form Ω = dp ∧ dq. The Hamiltonian function 1 2 is given by F (p, q) = 2 |p| . Solving the Hamiltonian system ½ p˙ = 0 q˙ = p

we find the Hamiltonian flow

ft(p, q) = (p, q + pt).

69 70 CHAPTER 8. GROWTH AND DYNAMICS

Let us describe the dynamics of this flow. Every torus {p = a} is in- variant under {ft}, and moreover the restriction of the flow to each such torus is simply the (quasi)-periodic motion q → q + at. All these tori are homologous to the zero section of the cotangent bundle. Ge- ometrically they correspond to families of parallel Euclidean lines on Tn. The flow above belongs to an important class of time-independent Hamiltonian systems called integrable systems (see [Ar]). Integrable systems can be characterized by the fact that their energy levels are foliated (up to measure zero) by invariant middle-dimensional tori which carry (quasi)-periodic motion. Traditionally this is considered as the simplest pattern of dynamical behaviour in classical mechanics. A natural question is what happens to the invariant tori when one per- turbs the system. The Kolmogorov-Arnold-Moser theory (or in short KAM theory, see [Ar]) tells us that for small perturbations most of the invariant tori persist. One has to require, however, that the rota- tion vector is “sufficiently irrational”. This means, in particular, that if in some angular coordinates θ on an invariant torus the dynamics ˙ n is given by θ = a, where a = (a1, ..., an) ∈ R , then a1, ..., an should be linearly independent over Q. For large perturbations the “topolog- ically essential” tori may disappear. For instance, one can deform the Euclidean metric on Tn to a Riemannian metric such that its geodesic flow has no invariant tori which carry quasi-periodic motion and are homologous to the zero section (see [AL]). Note that in our starting example the invariant tori {p = a} are Lagrangian. This is a general phenomenon which we are going to explain now. Exercise 8.1.A. Let F : M → R be a time independent Hamil- tonian on a symplectic manifold M. Consider a closed submanifold L ⊂ {H = c}. Show that if L is Lagrangian then L is invariant under the Hamiltonian flow of F . Hint: use linear algebra in order to prove that sgradF must be tangent to L. In some interesting cases the statement above can be reversed. Proposition 8.1.B. ([He]) Let F : T ∗Tn → R be a time independent Hamiltonian. Consider an invariant torus L ⊂ {H = c} carrying a 8.1. INVARIANT TORI OF CLASSICAL MECHANICS 71

quasi-periodic motion θ˙ = a where the coordinates of a are, as before, linearly independent over Q. Then L is Lagrangian.

Proof: Take a point x ∈ L, and assume that Ω| has the form P TxL bijdθi ∧ dθj. SinceP the dynamics is just a shift, θ(t) = θ(0) + at, we get that Ω|TyL = bijdθi ∧ dθj for every point y which lies on the trajectory of x. Note thatP every trajectory is dense on the torus, so we conclude that Ω = bijdθi ∧ dθj everywhere on L. But Ω|TL is an exact two-form. Hence bij = 0 for all i, j. Therefore Ω|TL = 0 and we have shown that L is Lagrangian. 2

Let F be a time-independent compactly supported Hamiltonian function on (T ∗Tn, Ω). Define a number E(F ) = sup |E|, where the supremum is taken over all E such that the energy level {F = E} contains a Lagrangian torus homologous to the zero section. How can one find a non-trivial upper bound for E(F )? Questions of this flavor 1 are studied in the framework of converse KAM theory. Write {ft} for the Hamiltonian flow generated by F .

Theorem 8.1.C. (cf. [BP2],[P8]) ρ(1l, ft) ≥ E(F )t for all t ∈ R.

Proof: Note that it suffices to prove the inequality for t = 1 (make a time reparametrization). Let L ⊂ {H = E} be a Lagrangian torus ho- mologous to the zero section. Then L has the stable Lagrangian inter- sections property (see 6.2.D). Since (T ∗Tn, Ω) is a Liouville manifold the length spectrum of Ham(M, Ω) equals {0} (see 7.3.B). Therefore all the assumptions of Theorem 7.4.A are satisfied. This theorem im- plies that ρ(1l, f1) ≥ E. Taking the supremum over all such E, we get the required estimate. 2

The geometric content of this estimate should be understood in a more general context of the growth of one-parameter subgroups of Hamiltonian diffeomorphisms.

1The goal of KAM theory is to prove existence of invariant Lagrangian tori, while converse KAM studies obstructions to their existence, see e.g. [Mac] and references therein. 72 CHAPTER 8. GROWTH AND DYNAMICS 8.2 Growth of one-parameter subgroups

Let (M, Ω) be a symplectic manifold, and let {ft} be a one-parameter subgroup of Ham(M, Ω) generated by a normalized Hamiltonian func- tion F ∈ A. One of the central problems of Hofer’s geometry is to ex- plore interrelations between the function ρ(1l, ft) and the dynamics of the flow {ft}. For instance, Theorem 8.1.C above states that invariant tori of an autonomous Hamiltonian flow on T ∗Tn which are homol- ogous to the zero section and carry quasi-periodic motion contribute to the linear growth of the function ρ(1l, ft). There exists another, purely geometrical reason for the interest in this function. It comes from the theory of geodesics of Hofer’s metric (see Chapter 12 below). A Hamiltonian path {ft} is called a strictly minimal geodesic if each of its segments minimizes the length between its end points. Conjec- turally (see 12.6.A below) all sufficiently short segments of any one- parameter subgroup are strictly minimal geodesics (in other words, every one-parameter subgroup is locally strictly minimal). However as we will see in 8.2.H long segments may lose the minimality. The min- imality breaking is an intriguing phenomenon which is still far from being understood. At present several approaches to this phenomenon are known. One of them, based on the theory of conjugate points in Hofer’s geometry is presented in Chapter 12. Here we discuss an ap- proach based on the notion of asymptotic growth of a one-parameter subgroup (see [BP2]). The asymptotic growth is defined as

ρ(1l, f ) µ(F ) = lim t . t→+∞ t||F ||

Exercise. Show that the limit above exists. Hint: Use that the func- tion ρ(1l, ft) is subadditive: ρ(1l, ft+s) ≤ ρ(1l, ft) + ρ(1l, fs). Clearly, µ(F ) always belongs to [0; 1]. If µ(F ) < 1 then the path {ft} is not a strictly minimal geodesic. Let us consider several examples of the behaviour of the func- tion ρ(1l, ft). It was proved by Hofer [H2] that every one-parameter subgroup of Ham(R2n) is locally strictly minimal. On the other hand Sikorav [S2] discovered the striking fact that each such subgroup re- 8.2. GROWTH OF ONE-PARAMETER SUBGROUPS 73

mains a bounded distance of identity, and thus cannot be globally strictly minimal.

2n Theorem 8.2.A. Let {ft} be a one-parameter subgroup of Ham(R ) generated by a compactly supported Hamiltonian function F . Assume that the support of F is contained in a Euclidean ball of radius r. Then 2 the function ρ(1l, ft) is bounded: ρ(1l, ft) ≤ 16πr .

We refer the reader to [HZ, p. 177] for the detailed proof of this theo- rem (see also discussion in 12.6.E below). Let us return now to one-parameter subgroups of Ham(T ∗Tn). First of all, it is proved in [LM2] that all of them are locally strictly minimal. In other words ρ(1l, ft) = t||F || provided t is small enough. Of course, this implies that our estimate 8.1.C above is not sharp for small t. Indeed, in general E(F ) is strictly smaller than ||F ||. Nevertheless, in the case when n = 1 and F ≥ 0 estimate 8.1.C is asymptotically sharp! Notice that since every closed curve on the cylinder T ∗T1 is Lagrangian, the quantity E(F ) is simply the supremum of those real numbers E for which the level {F = E} contains a non-contractible embedded circle.

Theorem 8.2.B. ([PS]) Let F be a non-negative compactly sup- ported Hamiltonian on the cylinder T ∗T1 with ||F || = 1. Then the converse KAM parameter E(F ) coincides with the asymptotic growth µ(F ): E(F ) = µ(F ).

Proof: We have to prove that µ(F ) ≤ E(F ). Combining this with 8.1.C we get the desired result. If E(F ) = max F = 1 then 8.1.C yields that µ(F ) = E(F ). Suppose now that E(F ) < 1. The idea is to decompose the flow {ft} into a product of two commuting flows with simple asymptotic behaviour. Choose ² > 0 small enough, and consider a smooth non-decreasing function u : [0; +∞) → [0; +∞) with the following properties: • u(s) = s for s ≤ E(F ) + ²;

• u(s) = E(F ) + 2² for s ≥ E(F ) + 3² ; 74 CHAPTER 8. GROWTH AND DYNAMICS

• u(s) ≤ s for all s.

Consider the new Hamiltonians G = u ◦ F and H = F − G, and denote by {gt} and {ht} the corresponding Hamiltonian flows. These flows commute and ft = gtht. Thus

ρ(1l, ft) ≤ ρ(1l, gt) + ρ(1l, ht). (8.2.C)

Note that ||G|| ≤ E(F ) + 2², thus

ρ(1l, gt) ≤ t(E(F ) + 2²). (8.2.D)

Further, the support of H is contained in a subset D² = {F ≥ E(F ) + ²}. For a generic sufficiently small ² the set D² is a domain whose boundary consists of contractible closed curves (we use here the definition of E(F )). Assume now that the support of F is contained in an annulus ¯ ¯ A = {(p, q) ∈ T ∗T1 ¯ |q| ≤ a/2} for some a > 0. Note that ∂D² ⊂ A. Hence the set D² is contained in some set D0 ⊂ A which is a finite union of pairwise disjoint closed discs of total area at most a. Since the cylinder has the infinite area, it is an easy consequence of the Dacorogna-Moser theorem [HZ, Sect. 1.6] that there exists a symplectic embedding i : R2 → T ∗T1, and a finite union D00 ⊂ R2 of Euclidean discs such that i maps D00 diffeomorphically on D0. Clearly, i induces the natural homomorphism

2 ∗ 1 i∗ : Ham(R ) → Ham(T T ).

It is important to notice that i∗ does not increase the corresponding Hofer distances. Our flow ht lies in the image of i∗, i.e., ht = i∗(et) 2 where et is a one–parameter subgroup of Ham(R ) whose Hamiltonian is supported in D00. Thus Theorem 8.2.A above implies that

ρ(1l, ht) ≤ 16a.

Combining this inequality with (8.2.D) and (8.2.C) we get that

ρ(1l, ft) ≤ t(E(F ) + 2²) + 16a (8.2.E) 8.2. GROWTH OF ONE-PARAMETER SUBGROUPS 75

for all t > 0. Dividing by t and passing to the limit when t → +∞ we get that µ(F ) ≤ E(F ) + 2². Since ² is arbitrary small, this completes the proof. 2

Remark 8.2.F. The same proof shows that if E(F ) = 0 then the function ρ(1l, ft) is bounded. Indeed, since (8.2.E) holds for all ² > 0 we get that ρ(1l, ft) ≤ 16a. In view of the inequality E(F ) ≤ µ(F ) this implies the following “rigidity”-type statement: if µ(F ) = 0 then ρ(1l, ft) is bounded (see 8.4 below for further discussion). Theorem 8.2.B and Remark 8.2.F hold true for all open surfaces of infinite area (see [PS]). Moreover, one can easily modify the definition of E(F ) and extend these statements to arbitrary (not necessarily non- negative) Hamiltonians F . At the moment, no generalization of 8.2.B to higher dimensions is known. The next discussion shows however that estimate 8.1.C is sharp at least in the following very special situation. Let F : T ∗Tn → R be a compactly supported time independent Hamiltonian function which satisfies the following conditions: (i) F ≥ 0 (ii) max F = 1 (iii) the maximum set Σ = {F = 1} is a smooth section of the cotangent bundle. It turns out that the geometry of the corresponding flow {ft} drastically depends on whether Σ is Lagrangian or not! Suppose that Σ is a Lagrangian submanifold. Then by definition E(F ) = 1. Therefore Theorem 8.1.C above implies that {ft} is a strictly minimal geodesic, and in particular E(F ) = µ(F ). Assume now that Σ is not Lagrangian, that is Ω does not vanish at least on one tangent space to Σ. In this case it is unknown whether µ(F ) equals E(F ) or not. However we claim that estimate 8.1.C is at least non-trivial, namely

E(F ) ≤ µ(F ) < 1. (8.2.G)

In order to explain this inequality, we need the following fairly general result which in some interesting situations allows us to show that µ(F ) is strictly less than 1. 76 CHAPTER 8. GROWTH AND DYNAMICS

Theorem 8.2.H. Let F be a time-independent normalized Hamilto- nian on a symplectic manifold (M, Ω). Let Σ+ and Σ− be the maximum and the minimum sets of F respectively. Suppose that there exists φ ∈ Ham(M, Ω) such that either φ(Σ+) ∩ Σ+ = ∅ or φ(Σ−) ∩ Σ− = ∅. Then µ(F ) < 1, and in particular the Hamiltonian flow of F is not a strictly minimal geodesic.

The proof is given in 8.3 below. Let us sketch the proof of 8.2.G. One can show that since Σ is not Lagrangian there exists a Hamiltonian diffeomorphism φ such that φ(Σ) ∩ Σ = ∅. Thus combining Theorems 8.1.C and 8.2.H we get that E(F ) ≤ µ(F ) < 1. The proof of the existence of the Hamiltonian diffeomorphism φ which displaces the non-Lagrangian submanifold Σ is complicated. It is based on Gromov’s h-principles for partial dif- ferential relations. In fact one can even show more, namely that the displacement energy of Σ vanishes. We refer the reader to [P2], [LS] for the proof of this result and its generalizations. Let us mention also that Siburg [Si2] extended inequality 8.1.C to time-dependent Hamiltonian flows. Note that in the time-depen- dent case it is already non-trivial to define the converse KAM-type parameter because of the lack of the energy conservation. Siburg’s definition is based on the concept of minimal action developed by Mather.

8.3 Curve shortening in Hofer’s geome- try

Let (M, Ω) be a symplectic manifold. For every function F ∈ A,F 6≡ 0 set kF + F ◦ φk δ(F ) = inf , φ 2kF k where the infimum is taken over all φ ∈ Ham(M, Ω). Theorem 8.3.A. ([BP2])

µ(F ) ≤ δ(F ). 8.3. CURVE SHORTENING IN HOFER’S GEOMETRY 77

Theorem 8.2.H above is an immediate consequence of this result. In- deed, if F satisfies the assumptions of 8.2.H then δ(F ) < 1. Proof of 8.3.A: Without loss of generality, assume that kF k = 1. Take φ ∈ Ham(M, Ω) and T > 0. Write ³ ´ ³ ´ −1 −1 −1 f2T = fT ◦ φ ◦ fT ◦ φ ◦ φ ◦ fT ◦ φ ◦ fT = A ◦ B.

Observe that B is a commutator so ρ(1l,B) ≤ 2ρ(1l, φ). The diffeo- −1 morphism A is generated by a path gt = ftφftφ for t ∈ [0,T ] and the corresponding Hamiltonian is

−1 −1 G(x, t) = F (x) + F (φ ft x).

So we have

−1 −1 kGtk = kF + F ◦ φ ◦ ft k −1 = kF ◦ ft + F ◦ φ k = kF + F ◦ φ−1k

since F ◦ ft = F (energy conservation). So

−1 ρ(1l, f2T ) ≤ T kF + F ◦ φ k + 2ρ(1l, φ)

and hence ρ(1l, f ) 1 ρ(1l, φ) 2T ≤ kF + F ◦ φ−1k + 2T 2 T for all φ ∈ Ham(M, Ω). Now let T → ∞ and deduce that µ(F ) ≤ δ(F ). 2

We refer the reader to [LM2] and [P9] for various curve shortening procedures in Hofer’s geometry. We will return to this issue in Chapter 11 below. The quantity δ(F ) admits the following natural generalization. Set PN−1 1 k j=0 F ◦ φjk δN (F ) = inf , φ1,...,φN−1 N kF k 78 CHAPTER 8. GROWTH AND DYNAMICS

where φ0 = 1l and the infimum is taken over all sequences {φj}, j = 1, ..., N−1 of Hamiltonian diffeomorphisms. With this language δ = δ2. One can show [P9] that on a closed symplectic manifold δN (F ) → 0 when N → +∞ for every function F ∈ A. I do not know the decay rate of the sequence δN (F ).

8.4 What happens when the asymptotic growth vanishes?

As we have seen in 8.2.F above, on the cylinder (and more gener- ally, on open surfaces of infinite area) every one-parameter subgroup whose asymptotic growth vanishes remains a bounded distance of the identity. What happens on other symplectic manifolds? Problem 8.4.A. Does there exist a symplectic manifold (M, Ω) and a one-parameter subgroup {ft} of Ham(M, Ω) such that the function ρ(1l, ft) has√ an intermediate asymptotics at infinity (for instance, it grows like t)? This problem is open even for such an elementary symplectic manifold as the 2-torus T2. The only excuse is that the functions F ∈ A(T2) with µ(F ) = 0 are non-generic, as the next result shows.

Theorem 8.4.B. Assume that 0 is a regular value of F ∈ A(T2). Then µ(F ) > 0.

Proof: The set D = {F = 0} consists of a finite number of pairwise disjoint embedded circles. Thus there exists a non-contractible simple closed curve L ⊂ T2 such that L∩D = ∅. Thus for some C > 0 we have 2 that |F (x)| > C for all x ∈ L. Since π1(Ham(T )) = 0 (see 7.2.B), it follows from 7.4.A that ρ(1l, ft) ≥ Ct for all t, and in particular µ(F ) > 0. 2 Chapter 9

Length spectrum

In this chapter we describe a method of calculation of the length spec- trum in Hofer’s geometry. Our approach is based on the theory of symplectic fibrations over the 2-sphere.

9.1 The positive and negative parts of Hofer’s norm

Let (M, Ω) be a closed symplectic manifold. For γ ∈ π1(Ham(M, Ω)) we set

Z 1 ν+(γ) = inf max F (x, t)dt = inf max F (x, t), F 0 x F x,t Z 1 ν−(γ) = inf − min F (x, t)dt = inf (− min F (x, t)), F 0 x F x,t where the infimum is taken over all normalized periodic Hamiltonians F ∈ H which generate a loop representing γ. The correctness of this definition can be proved exactly in the same way as Lemma 5.1.C above.

Exercise. Prove that ν+(γ) = ν−(−γ), and ν(γ) ≥ ν−(γ) + ν+(γ) (cf. the open problem in 2.4 above).

79 80 CHAPTER 9. LENGTH SPECTRUM

The calculation of ν+(γ) turns out to be non-trivial evenR in the 2 following simplest case. Consider (S , Ω) normalized so that S2 Ω = 1. Let γ be the class of the 1-turn rotation {ft}, and let F ∈ H be the Hamiltonian generating {ft}. We have seen in 6.3.C above that 1 max F = − min F = 2 . (The factor 2π in 6.3.C disappeared in view of 1 the normalization of Ω above.) Thus ν+(γ) ≤ 2 . In fact equality holds! 1 Theorem 9.1.A. ν+(γ) = 2 .

A Hamiltonian loop {ft} representing a class γ 6= 0 is called a closed minimal geodesic if length{ft} = ν(γ). Note that a closed minimal geodesic is never strictly minimal.

Corollary. ([LM2]) ν(γ) = 1 and {ft} is a closed minimal geodesic. 1 1 Proof: First of all, length{ft} = 2 − (− 2 ) = 1 and hence ν(γ) ≤ 1. Since 2γ = 0 we deduce that

1 ν (γ) = ν (−γ) = ν (γ) = . − + + 2

Furthermore, ν(γ) ≥ ν−(γ) + ν+(γ) = 1. We conclude that ν(γ) = 1 and {ft} is a minimal geodesic. 2

Theorem 9.1.A is proved in 9.4 below. We refer to [P3] for gener- alizations to CP n, n ≥ 2.

9.2 Symplectic fibrations over S2

Let (M, Ω) be a closed symplectic manifold. We will assume from now on 1 that H1(M, R) = 0. As a consequence Ham(M, Ω) coincides with the connected component of 1l in Symp(M, Ω). Let p : P → S2 be a smooth fibration with fiber M endowed with a fiberwise symplectic structure as follows. For each x ∈ S2, p−1(x) is endowed with a sym- −1 plectic form Ωx such that Ωx depends smoothly on x and (p (x), Ωx)

1It is not hard to get rid of this assumption, see the theory of Hamiltonian symplectic fibrations in [MS] and [P4]. 9.2. SYMPLECTIC FIBRATIONS OVER S2 81

is symplectomorphic to (M, Ω). In addition, we always choose an ori- entation on S2 as part of the data (so P is also oriented). We call p : P → S2 a symplectic fibration (for more details see [MS]). Every loop {ft} of Hamiltonian diffeomorphisms of M gives rise 2 2 to a symplectic fibration. Take two copies of the unit 2-disc D+,D− 2 2 such that D+ is endowed with the positive orientation and D− with the reversed one. Define a new manifold

2 2 P = M × D− ∪ψ M × D+ where ψ is the gluing map

1 1 ψ : M × S → M × S , (z, t) 7→ (ftz, t). Clearly P has the natural structure of a symplectic fibration over 2 2 S since the ft are symplectomorphisms (and moreover S gets an orientation in the result of this construction). Homotopic loops lead to isomorphic symplectic fibrations i.e. there exist smooth preserving the fiberwise symplectic struc- ture and the orientation. Furthermore, the construction can be re- versed. Given a fibration P → S2 together with a trivialization over one point, one can reconstruct the homotopy class γ. Note also that the class γ = 0 corresponds to the trivial fibration S2 × (M, Ω). We leave the proof of these statements to the reader. We shall write P = P (γ), where γ is the homotopy class of {ft}. Let us visualize the fibration P (γ) for the class of the 1-turn rotation of S2. Note that in this case both the base and the fiber are S2. It will be useful to identify S2 with the complex projective line CP 1 and move to the complex setting. At this point we will make a digression concerning the of complex projective spaces. Let E be a 2n-dimensional real vector space endowed with a complex structure j (here j is a linear transformation E → E with j2 = −1l), a scalar product g and a symplectic form ω such that g(ξ, η) = ω(ξ, jη) for all ξ, η ∈ E (cf. 4.1 above). Of course, using j we can consider E as a complex vector space. The pair (ω, g) as above is called a Hermitian 82 CHAPTER 9. LENGTH SPECTRUM

structure on the complex space (E, j). Take the unit sphere S = {ξ ∈ E | g(ξ, ξ) = 1}. Consider the circle action on S defined by ξ → e2πjtξ, t ∈ R/Z. The orbits of this action are precisely the sets of the form S ∩ l, where l is a complex line in E. Thus the space of orbits S/S1 can be canonically identified with the complex projective space P(E). The action preserves the restriction of ω to TS. Denote by Ω the projection 1 of π ω to P(E). This form is closed. Further, an elementary linear algebra argument shows that Ω is non-degenerate, and thus we get a symplectic form on P(E). The form Ω is called the standard (or the Fubini-Study) sym- plectic form on P(E) associated to the Hermitian structure (ω, g) on a complex space (E, j). The construction above is a particular case of the Marsden-Weinstein reduction [MS] which plays a fundamental role in the theory of group actions on symplectic manifolds. The factor 1 π above is chosen for the following reason. Exercise 9.2.A. Show that the integral of Ω over a projective line in P(E) equals 1. It follows from Moser’s theorem [MS] that different Hermitian struc- tures on (E, j) give rise to diffeomorphic symplectic forms on P(E). Exercise 9.2.B. Consider the space Cn endowed with the standard Hermitian structure (see 4.1). Show that the standard symplectic form on P(Cn) = CP n−1 is invariant under the group PU(n) = U(n)/S1, and that this group acts on CP n−1 by Hamiltonian diffeomorphisms. Show that PU(2) = SU(2)/{1l; −1l}. Exercise 9.2.C. Consider a loop of projective unitary transforma- 1 1 tions of CP which in homogeneous coordinates (z1 : z2) on CP is given by −2πit (z1 : z2) → (e z1 : z2), t ∈ [0; 1]. 1 Show that this loop represents the non-trivial element of π1(Ham(CP )). Hint: Use the canonical isomorphism SU(2)/{1l; −1l} → SO(3) (see [DFN]). 9.3. SYMPLECTIC CONNECTIONS 83

There exists a natural “parametric” version of the construction above which allows us to produce symplectic fibrations whose fiber is CP n−1 endowed with the standard symplectic form. Let E → S2 be a complex of rank n, and let P(E) be its projectivization. Every Hermitian structure on E gives rise to the fiber-wise symplectic form Ωx on P(E). Here Ωx equals the standard symplectic form on the fiber P(Ex). Different choices of the Hermitian structure lead to isomorphic symplectic fibrations. We return now to the symplectic fibration P (γ), where γ is the 2 1 non-trivial element of π1(Ham(S )). Let T → CP be the tautological line bundle i.e. the one whose fiber over a complex line in C2 is the line itself. Let C = C × CP 1 be the trivial bundle. Exercise 9.2.D. Prove that the symplectic fibration P (γ) is isomor- phic to P(T ⊕ C). Hint: Represent the base CP 1 as the union of two discs ¯ 1 ¯ D− = {(x0 : x1) ∈ CP ¯ |x0/x1| ≤ 1}, and ¯ 1 ¯ D+ = {(x0 : x1) ∈ CP ¯ |x1/x0| ≤ 1}.

1 Introduce a coordinate t on the circle S = ∂D+ = ∂D− by x1/x0 = 2πit e . The bundle T ⊕ C can be trivialized over D− and D+ in such a way that the transition function S1 × C2 → S1 × C2 has the form

−2πit (t, z1, z2) → (t, e z1, z2) (see e.g. [GH] for the detailed description of the tautological line bun- dle). The result now follows from 9.2.C above. Exercise 9.2.E. Show that P(T ⊕ C) is biholomorphically equivalent to the complex blow up of CP 2 at one point. The fibration is obtained by the proper transform of the pencil of lines passing through the blown up point.

9.3 Symplectic connections

Let p : P → S2 be a symplectic fibration with fiber (M, Ω). A connec- tion σ on P (that is a field of 2-dimensional subspaces transversal to 84 CHAPTER 9. LENGTH SPECTRUM

the fibers, see figure 9), is called symplectic if the parallel transport preserves the fiberwise symplectic structure. One can show that every symplectic fibration admits a symplectic connection, see [GLS],[MS].

P

p

2 S Figure 9

Example. Let E → S2 be a complex vector bundle endowed with a Hermitian metric. Every Hermitian connection on E induces a sym- lectic connection on the projectivized bundle P(E) → S2. Recall the definition of curvature of a connection. For x ∈ S2 and 2 ξ, η ∈ TxS , extend ξ and η locally in a neighbourhood of x and con- sider horizontal lifts ξ,˜ η˜ of the obtained vector fields. By definition, the curvature ρσ at the point x is a 2-form which takes values in the Lie ¡ ¢vert algebra of vector fields on the fiber p−1(x). So ρσ(ξ, η) = [ξ,˜ η˜] , where ’vert’ denotes the projection of [ξ,˜ η˜] onto the fiber p−1(x). Thus if σ is a symplectic connection, then ρσ(ξ, η) lies in the Lie algebra of Symp(p−1(x)), that is ρσ(ξ, η) is a Hamiltonian vector field since H1(M, R) = 0. Identifying the Hamiltonian vector fields with the nor- malized Hamiltonian functions, we shall consider ρσ(ξ, η) as a normal- ized function p−1(x) → R. R 2 Fix an area form τ on S such that S2 τ = 1 (here we need the orientation of S2). Since every 2-form on S2 is a multiple of τ we can write ρσ = Lστ, where Lσ is a function on P . 9.3. SYMPLECTIC CONNECTIONS 85

The theory of symplectic connections has been recently developed by Guillemin, Lerman and Sternberg (see [GLS], [MS]). As funda- mental objects of this theory arise the coupling form of a symplectic connection and the coupling class of a symplectic fibration. For a point (x, z) ∈ P we can write

−1 2 T(x,z)P = Tzp (x) ⊕ TxS ,

which is the decomposition with respect to the symplectic connection σ. Define the coupling form δσ of the connection σ as the 2-form on P given by

σ σ δ (v ⊕ ξ, w ⊕ η) = Ωx(v, w) − ρ (ξ, η)(z).

−1 −1 2 σ Here z ∈ p (x), v, w ∈ Tzp (x) and ξ, η ∈ TxS , while ρ (ξ, η) is considered as a function on the fiber p−1(x). It turns out that the coupling form is closed. Denote by c its co- homology class in H2(P, R). Obviously the restriction of c to any fiber −1 p (x) coincides with the class [Ωx] of the symplectic form. Further- more, one can prove that cn+1 = 0 where 2n = dimM. The next result shows that the class c is uniquely determined by these two properties.

Theorem 9.3.A. (See [GLS], [MS]) The class c is the unique coho- 2 n+1 mology class in H (P, R) such that c|fiber = [Ωx] and c = 0. In particular, c is an invariant of the symplectic fibration P which does not depend on the choice of the connection σ. We call c the coupling class of P . Detailed proofs of all these results can be found in [GLS],[MS]. The next construction plays an important role in our approach to the length spectrum.

The weak coupling construction. ([GLS],[MS]) For ε > 0 suffi- ciently small there exists a smooth family of closed 2-forms ωt on P with t ∈ [0, ε) such that

∗ • ω0 = p τ 86 CHAPTER 9. LENGTH SPECTRUM

∗ • [ωt] = tc + p [τ]

• ωt|fiber = tΩx

• ωt is symplectic for all t > 0. Set ε(P ) = sup ε, where the supremum is taken over all such deformations. This quantity measures how strong the weak coupling can be. Note that for a trivial fibration, P = M × S2, ε(P ) = +∞. Thus in a sense ε(P ) measures the non-triviality of the fibration. There is another way to measure the non-triviality, which works for fibrations with other structure groups as well. It was suggested by Gromov [] for unitary vector bundles and we will here give a symplectic version. The idea is to measure the minimal possible norm of the curvature of a symplectic connection on P . We will introduce the following concept. Definition. 1 χ+(P ) = sup σ σ max L P is the (positive part of the) symplectic K-area of P . Here the supremum is taken over all symplectic connections on P , and Lσ is defined by ρσ = Lστ. Exercise. Prove that both quantities introduced above, ε(P ) and χ+(P ), do notR depend on the particular choice of the area form τ 2 on S with S2 τ = 1. Hint: First apply Moser’s theorem [MS] which states that given any two such forms, say τ1 and τ2, there exists a 2 ∗ diffeomorphism a of S isotopic to the identity such that a τ2 = τ1. Then lift a to a fiber-wise symplectic diffeomorphism A of P , which means that p(A(z)) = a(p(z)) for all z ∈ P and the restriction Ax −1 ∗ of A to any fiber p (x) satisfies (Ax) Ωa(x) = Ωx. Such a lift can be constructed with the help of any symplectic connection on P . Note that A takes any weak coupling deformation associated to τ2 to a weak coupling deformation associated to τ1. This proves that ε(P ) does not depend on the choice of the area form. Further, A acts on the space of symplectic connections on P by σ → A∗σ. Show that

A∗σ σ ρ (a∗ξ, a∗η)(Az) = ρ (ξ, η)(z) 9.3. SYMPLECTIC CONNECTIONS 87

2 for every point z ∈ P and every pair of vectors ξ, η ∈ Tp(z)S . The fact that χ+(P ) does not depend on the choice of the area form is a simple consequence of this formula.

1 Theorem 9.3.B. ([P4]) Let P = P (γ). Then ε(P ) = χ+(P ) = . ν+(γ)

We will prove a weaker statement, namely that ε(P ) ≥ χ+(P ) ≥ 1 , but this will be enough to prove Theorem 9.1.A. ν+(γ)

Proof of ε(P ) ≥ χ+(P ): Let σ be a symplectic connection. Consider

∗ σ ωt = p τ + tδ ,

where δσ is the coupling form. At the point (x, z) ∈ P, we have

σ σ ωt,(x,z) = tΩx ⊕ −tL (x, z)τ + τ = tΩx ⊕ (1 − tL (x, z))τ.

Clearly, ωt satisfies the first three properties of the weak coupling σ construction. The form ωt is symplectic when 1 − tL (x, z) > 0 or σ 1 equivalently L (x, z) < t for all x, z. This condition means that max Lσ(x, z) < 1 , or P t 1 > t. max Lσ(x, z) P Choose now an arbitrary κ > 0 and a symplectic connection σ such that 1 > χ (P ) − κ. maxLσ(x, z) + P

Thus there exists a coupling deformation ωt for t ∈ [0, χ+(P ) − κ). That is for all κ > 0, ε(P ) ≥ χ+(P ) − κ and we conclude that

ε(P ) ≥ χ+(P ).

2

Note, that we proved the existence of the weak coupling deforma- tion. 1 Proof of χ+(P ) ≥ : The proof is based on the following exercise. ν+(γ) 88 CHAPTER 9. LENGTH SPECTRUM

Exercise. Let p : P → S2 be a symplectic fibration. Let ω be a closed 2-form on P such that ω|fiber = Ωx. Set −1 σ(x,z) = {ξ ∈ T(x,z)P | iξω = 0 on Tzp (x)}. Show that σ defines a symplectic connection on P.

Let {ft}, t ∈ [0, 1], be an arbitrary loop of Hamiltonian diffeo- morphisms generated by a normalized Hamiltonian F ∈ H. Fix polar coordinates u ∈ (0, 1] (the radius) and t ∈ S1 = R/Z (the normal- ized angle) on D2. Take a monotone cut off function φ(u) such that φ(u) = 0 near u = 0 and φ(u) = 1 near u = 1. Set

2 2 P = M × D− ∪ψ M × D+, where ψ(z, t) = (ftz, t). Define a closed 2-form ω on P by ½ 2 Ω on M × D+ ω = 2 Ω + d(φ(u)Ht(z)) ∧ dt on M × D− where Ht(z) = F (ftz, t). Exercise. Prove that ω is well defined i.e. show that ψ∗Ω = Ω + dHt ∧ dt ( this can be done by a direct computation). Let us compute the curvature of ρσ of the symplectic connection σ 2 associated to ω. Notice that ρ vanishes on D+ (since Ω induces a flat 2 σ 2 connection on D+), so it remains to compute ρ on D−. Note that σ 2 ρ = 0 near 0 on D−, which is good since we avoid the singularity at zero. In order to compute the curvature, we must find the horizontal ∂ ∂ 2 lifts of ∂u and ∂t at (x, z) ∈ M × D−. ∂ Let the horizontal lift of ∂u be of the form

f∂ ∂ = + v for some v ∈ T p−1(x). ∂u ∂u z

f∂ Since by definition of the connection, we have ω( ∂u , w) = 0 for all w ∈ TzM, we have

³ f∂ ´ 0 = ω , w = ω(v, w) = Ω(v, w) ∂u 9.3. SYMPLECTIC CONNECTIONS 89

−1 for all w ∈ Tzp (x). It follows from the non-degeneracy of Ω that v = 0 and hence that f∂ ∂ = . ∂u ∂u Setting f∂ ∂ = + v for some other v ∈ T p−1(x), ∂t ∂t z −1 we get as before, that for all w ∈ Tzp (x) ³f∂ ´ ∂ 0 = ω , w = ω( + v, w) = Ω(v, w) − d(φ(u)H )(w). ∂t ∂t t

So ivΩ = d(φ(u)Ht) and therefore v = − sgrad φ(u)Ht = −φ(u) sgrad Ht and we can deduce that f∂ ∂ = − φ(u) sgrad H . ∂t ∂t t Computing the curvature we get ³ ∂ ∂ ´ h ∂ ∂ ivert ρν , = − φ(u) sgrad H , = φ0(u) sgrad H ∂t ∂u ∂t t ∂u t (actually the commutator is already a vertical vector field). Passing from vector fields to Hamiltonian functions in the definition of curva- ture we see that ³ ∂ ∂ ´ ρσ , = φ0(u)H (z). ∂t ∂u t 2 Fix κ > 0 and choose the area form on D− to be (1 − κ)dt ∧ du 2 (remember that D− has the reversed orientation). 2 2 Extend it to an area form on D+ such that Area (D+) = κ. Denote the resulting area form by τ. We see that ρσ = Lσ(u, t, z)τ, where

½ 2 σ 0 on M × D+ 0 L (u, t, z) = φ (u)Ht(z) 2 1−κ on M × D−. 0 Finally, choose φ such that φ (u) ≤ 1+κ and {ft} such that max Ft = z max Ht ≤ ν+(γ) + κ. We have z

σ 1 + κ max L ≤ (ν+(γ) + κ) P 1 − κ 90 CHAPTER 9. LENGTH SPECTRUM

so 1 1 − κ χ+(P ) = sup σ ≥ . σ max L (1 + κ)(ν+(γ) + κ) P Since κ was arbitrary we have proved that 1 χ+(P ) ≥ . ν+(γ) 2

9.4 An application to length spectrum

The final step in the proof of Theorem 9.1.A is the following estimate for ε(P ) which we will discuss in the next chapter. Theorem 9.4.A. Let P = P (γ), where γ is the 1-turn rotation of S2. Then ε(P ) ≤ 2.

1 Proof of 9.1.A: We have seen that ν+(γ) ≤ 2 . On the other hand, 9.3.B and 9.4.A imply 1 2 ≥ ε(P ) ≥ χ+(P ) ≥ , ν+(γ) 1 so ν+(γ) = 2 and ε(P ) = χ+(P ) = 2. 2 Exercise. Let C and T be the trivial and the tautological holomorphic 1 line bundles over CP respectively (see 9.2). Let ∇C be the natural flat connection on C. There exists a unique connection, say ∇T , on T which preserves the canonical Hermitian structure on T coming from C2, and such that its (0, 1)-part coincides with the ∂¯-operator (see [GH]). Consider symplectic fibration P = P(T ⊕ C) and denote by σ the connection on P which comes from ∇T ⊕ ∇C . Obviously, σ is symplectic (actually, parallel transport preserves both the symplectic and the complex structure on the fibers). Let τ be the Fubini-Study area form on CP 1 defined in 9.2 above. Prove that 1 = 2, max Lσ P 9.4. AN APPLICATION TO LENGTH SPECTRUM 91 where ρσ = Lστ. In particular, σ is a connection with minimal possible curvature, where the ”size” of the curvature is measured with the help of the Fubini-Study area form τ. 92 CHAPTER 9. LENGTH SPECTRUM Chapter 10

Deformations of symplectic forms and pseudo-holomorphic curves

In the present chapter we prove the upper bound 9.4.A for the cou- pling parameter and therefore complete the calculation of the length spectrum of Ham(S2). This turns out to be a particular case of a more general problem concerning deformations of symplectic forms (see [P7]). We discuss an approach to the deformation problem which is based on Gromov’s theory of pseudo-holomorphic curves.

10.1 The deformation problem

Let (P, ω) be a closed symplectic manifold and l a ray in H2(P, R) with origin at [ω], see figure 10.

Problem. How far can one deform ω through symplectic structures on P in such a way that the cohomology class moves along l? Note that adding a small closed 2-form to ω leaves it symplectic, but how far can one go? In the case of the coupling deformation, the form [p∗τ] itself is degenerate, but for small t,[p∗τ] + tc is already

93 94 CHAPTER 10. DEFORMATIONS OF SYMPLECTIC FORMS

H2(P, R) l

. [ω]

Figure 10

symplectic. In this situation we can go as far along the ray in the direction of the coupling class c as ε(P ). Here is an example of an obstruction to going infinitely far. As- sume that dim P = 4 and that Σ ⊂ P is an embedded 2-sphere such that ω|T Σ is an area form (in other words Σ is a symplectic submani- fold of P ). We write (A, B) for the intersection index of two homology classes A and B. Definition. Let Σ ⊂ P 4 be a symplectic embedded sphere. If ([Σ], [Σ]) = −1 we call Σ an exceptional sphere.

Theorem 10.1.A. ([McD1]) Let Σ ⊂ (P 4, ω) be an exceptional sphere. Let ωt, t ∈ [0, 1] be a deformation of ω through symplectic forms. Then ([ω1], [Σ]) > 0. In other words, the hyperplane (x, [Σ]) = 0 in H2(P, R) is a wall which cannot be crossed by a symplectic deformation. The proof is based on the theory of pseudo-holomorphic curves. Below we give a sketch of the proof of this statement as well as an application to the proof of 9.4.A.

10.2 The ∂¯-equation revisited

An almost complex structure j on a manifold P is a field of endomor- phisms TP → TP such that j2 = −1l. An important class of exam- ples comes from complex algebraic geometry. Every complex manifold (that is a manifold equipped with an atlas whose transition func- tions are holomorphic) has the canonical almost complex structure 10.2. THE ∂¯-EQUATION REVISITED 95 √ ξ → −1ξ. Almost complex structures arising in this way are called integrable. A deep fact is that integrability is equivalent to the van- ishing of a certain tensor associated to the almost complex structure ([NN]). As a consequence of this result one gets that every almost complex structure on a (real) surface is integrable. Moreover, on the 2-sphere all almost complex structures are diffeomorphic (a classical fact called the uniformization theorem [AS]). Definition. Let (P, ω) be a symplectic manifold. An almost complex structure j is compatible with ω if g(ξ, η) = ω(ξ, jη) defines a Rieman- nian metric on P . Exercise 10.2.A. Let (P, ω) be a real surface and let j be an almost complex structure on P . Then either j or −j is compatible with ω. Let (P, ω) be a symplectic manifold, and let j be a compatible in- tegrable almost complex structure on P . The triple (P, ω, j) is called a K¨ahler structure. For instance, CP n endowed with the standard ω and j is a K¨ahlermanifold (cf. 9.2 above). Thus every complex submanifold of CP n is K¨ahlerwith respect to the induced structure. Vice versa, if (P, ω, j) is a closed K¨ahlermanifold such that the cohomology class [ω] is integral, then (P, j) admits a holomorphic embedding into CP n for some n (this is the famous Kodaira embedding theorem, see [GH]). It was Gromov’s great insight [G1] that one can generalize some important methods in algebraic geometry to quasi-K¨ahlermanifolds (P, ω, j). Here j is a compatible almost complex structure but in general not integrable anymore. Remarkably enough, the theory of holomorphic curves extends without essential changes to the non- integrable case. The importance of such a generalization is due to the following reason. Every symplectic manifold admits a compati- ble almost complex structure. However there are symplectic manifolds which cannot have a K¨ahlerstructure (see [MS]). Exercise 10.2.B(see [MS]). Let E be an even dimensional linear space endowed with a non-degenerate skew-symmetric bilinear form ω. Show that the space of complex structures j : E → E, j2 = −1l is contractible. Therefore a compatible almost complex structure on a symplectic manifold is a section of a fiber bundle whose fibers are contractible. 96 CHAPTER 10. DEFORMATIONS OF SYMPLECTIC FORMS

Hence such structures do exist and moreover form a contractible space. By the same reason various extension problems connected to almost complex structures admit a positive solution. Exercise 10.2.C. Let (P, ω) be a 4-dimensional symplectic mani- fold, and let Σ ⊂ P be a symplectic submanifold. Assume that Σ is equipped with an almost complex structure j which is compatible with ω|T Σ. Show that j extends to a compatible almost complex structure on P . Let us turn now to the theory of pseudo-holomorphic curves on quasi-K¨ahlermanifolds. Definition. A map φ :(S2, i) → (P, j) is a pseudo-holomorphic (or j-holomorphic) curve if φ∗ ◦ i = j ◦ φ∗. Exercise 10.2.D. Show that on a complex manifold the definition above is equivalent to the usual Cauchy-Riemann equation. We set 1 ∂φ¯ = (φ + j ◦ φ ◦ i). 2 ∗ ∗

Exercise 10.2.E(cf. 4.1.A). Given (P, ω, j, g) and a j-holomorphicR curve φ :(S2, i) → (P, j), show that Area (φ(S2)) = φ∗ω. In R g S2 ∗ particular, if φ is non-constant then S2 φ ω > 0. Moreover, if φ is an embedding, then φ(S2) is a symplectic submanifold of P . Indeed, the restriction of the symplectic form coincides with the Riemannian area form. Let (P, ω) be a symplectic manifold. Choose an almost complex structure j compatible with ω. Then the TP gets the structure of a complex vector bundle. Since the space of compatible j’s is connected the corresponding characteristic classes do not depend on the choice of j. We write c1 for the first Chern class of TP with respect to any compatible almost complex structure. Exercise 10.2.F (Adjunction formula). Let (P 4, ω) be a symplectic manifold with a compatible almost complex structure j. Let Σ ⊂ P be an embedded j-holomorphic sphere. Show that 1 1 + (([Σ], [Σ]) − c (Σ)) = 0. 2 1 10.3. AN APPLICATION TO COUPLING 97

Hint: Use that ([Σ], [Σ]) is the self-intersection in the complex normal bundle νΣ and TΣP = T Σ ⊕ νΣ.

Corollary. For a symplectically embedded sphere Σ, c1 = 1 if and only if ([Σ], [Σ]) = −1. Proof: Since Σ is symplectically embedded, there exists an ω-compatible almost complex structure j such that Σ is j-holomorphic (see 10.2.C). The statement now follows from the adjunction formula 10.2.F. 2

10.3 An application to coupling

In this section we deduce 9.4.A from 10.1.A. Recall that we study the coupling deformation for the fibration P(T ⊕ C) → CP 1, where T and C are the tautological and trivial bundle respectively. Exercise. Let E be a complex vector space and let l ∈ P(E) be a line in E. Show that TlP(E) is canonically isomorphic to Hom(l, E/l) = l∗ ⊗ E/l. First of all, we wish to compute the (co)homology ring of P = P(T ⊕ C). Denote by [F ] the homology class of the fiber and by Σ the section corresponding to the rank 1 subbundle 0 ⊕ C. Clearly, ([F ], [F ]) = 0 and ([F ], [Σ]) = 1. In order to compute ([Σ], [Σ]), notice that the normal bundle νΣ is just the restriction on Σ of the tangent bundle to the fibers. The above exercise implies that

∗ νΣ = Hom(C,T ⊕ C/C) = Hom(C,T ) = C ⊗ T = T.

Thus c1(νΣ) = −1. We conclude that ([Σ], [Σ]) = −1 since([Σ], [Σ]) is just the self-intersection in the normal bundle. Let ωt be a coupling deformation. It is an easy consequence of Moser’s theorem [MS] that for small enough t, ωt is symplectomorphic to a K¨ahlerform with respect to the standard complex structure on P(T ⊕C). Without loss of generality, we may assume that ωt is a K¨ahler form for small t. Since Σ is a holomorphic section of P, we get that Σ is symplectic and embedded. Taking into account that ([Σ], [Σ]) = −1 98 CHAPTER 10. DEFORMATIONS OF SYMPLECTIC FORMS

we see that Σ is an exceptional sphere. Thus Theorem 10.1.A implies that ([ωt], [Σ]) > 0 for all t. ∗ 2 2 Recall, that [ωt] = p [τ]+tc, where [τ] is a generator of H (S , Z) coming from the orientation and c is the coupling class. The classes p∗[τ] and c are uniquely defined by the following relations: (p∗[τ], [F ]) = 0 (p∗[τ], [Σ]) = 1 (c, [F ]) = 1 c2 = 0. We shall use Poincar´eduality to identify homology and cohomol- ogy. ∗ 1 Exercise. Prove that p [τ] = [F ] and c = [Σ] + 2 [F ]. Thus, in view of Theorem 10.1.A t t t ([ω ], [Σ]) = ([F ] + t[Σ] + [F ], [Σ]) = 1 − t + = 1 − > 0, t 2 2 2 which implies that t < 2. Since this is true for any coupling deforma- tion, we have proved that ε(P ) ≤ 2. This completes the proof. 2

10.4 Some properties of pseudo-holomorphic curves in symplectic manifolds

Here we outline Gromov’s theory of pseudo-holomorphic curves (see 2n [G1], [AL]). Let (P , ω) be a symplectic manifold and let A ∈ H2(P, Z) be a primitive class. This means that A cannot be represented as kB where k > 1 is an integer and B ∈ H2(P, Z). In particular, A 6= 0. Let J be the space of all ω-compatible almost complex structures on P and let N be the space of all smooth maps f : S2 → P such that [f] = A. Define X ⊂ N × J by ¯ X = {(f, j)|f ∈ N , j ∈ J and ∂jf = 0}. Then X is a smooth submanifold of N × J and the projection π : X → J is a Fredholm operator i.e. Ker π∗ and Coker π∗ are finite dimensional. Its index satisfies

Index π∗ := dim (Ker π∗) − dim (Coker π∗) = 2(c1(A) + n). 10.4. PSEUDO-HOLOMORPHIC CURVES 99

The Fredholm property of π allows us to apply an infinite dimensional version of Sard’s theorem [Sm].

X

π

J γ j1 ..j0

Figure 11

Let j0 and j1 be regular values of π (i.e π∗ is surjective for all x ∈ −1 −1 −1 π (jk), k = 0, 1). So π (j0) and π (j1) are smooth submanifolds. −1 Then for a generic path γ joining j0 to j1, π (γ) is a smooth submani- −1 −1 −1 fold of dimension Index (π) + 1 and ∂π (γ) = π (j0) ∪ π (j1) (see figure 11). The crucial role is played by the compactness properties of π−1(γ). First of all, notice that π−1(γ) cannot be compact itself since it ad- mits the action of the non-compact group PSL(2, C). Here PSL(2, C) is the group of conformal transformations of (S2, i) and it acts as 2 2 follows. If h : S → S is a conformal transformation and (f, j0) ∈ −1 −1 π (j0) then we also have (f ◦ h, j0) ∈ π (j0). Exercise. Show that this action is free (use that A is primitive). Gromov’s compactness theorem. Either the moduli space −1 −1 π (γ)/P SL(2, C) is compact or there exists a family (fk, jk) ∈ π (γ) such that jk → j∞ and fk “converges” to a j∞-holomorphic cusp curve in the class A. 100 CHAPTER 10. DEFORMATIONS OF SYMPLECTIC FORMS

We do not specify the convergence here. What is important to us is that if π−1(γ)/P SL(2, C) is not compact, then there exists a j- holomorphic cusp curve in the class A for some j ∈ J . A cusp curve is defined as follows. Let A = A1 + ... + Ad for d > 1 be a decomposition 2 of A such that Ak 6= 0 for all k and let φk : S → P be j-holomophic curves in the classes Ak, k = 1, ..., d. SWe say that these data define a cusp curve in the class A. The union φk(Ak) is called its image and is usually assumed to be connected. Imagine now the following situation. • The set J 0 of those j ∈ J , which have cusp curves in the class A, has codimension at least 2 in J .

0 −1 • j0 ∈ J \J is a regular value of π and π (j0)/P SL(2, C) (which is thus a compact manifold without boundary) is not cobordant to zero (i.e. it does not bound a compact manifold). For instance, a point is not cobordant to zero, since it does not bound a compact manifold. 0 −1 Then for every regular value j1 ∈ J \J the set π (j1)/P SL(2, C) is not empty. Indeed, if it were empty, and we joined j0 to j1 by a generic path γ ∈ J \ J 0, we would have that

¡ −1 ¢ −1 ∂ π (γ)/P SL(2, C) = π (j0)/P SL(2, C).

−1 This contradicts the fact that π (j0)/P SL(2, C) is not cobordant to zero, see figure 12 which illustrates the contradiction.

This reflects the persistence principle, which we already encoun- tered in the case of pseudo-holomorphic discs in 4.2. Either solutions of the ∂¯-equation persist or bubbling off takes place.

10.5 Persistence of exceptional spheres

Here we sketch the proof of Theorem 10.1.A. Let (P 4, ω) be a sym- plectic manifold and let Σ ⊂ P be an exceptional sphere with A = [Σ]. Let ωt, t ∈ [0, 1] be a deformation of the symplectic form ω = ω0. 10.5. PERSISTENCE OF EXCEPTIONAL SPHERES 101

π

J j0 .. j J 0 1

Figure 12

Step 1. Choose an ω-compatible almost complex structure j0 such that Σ is j0-holomorphic and extend j0 to a one-parameter family jt such that jt is ωt-compatible (cf. 10.2.C above). Step 2. The theory described in the previous section holds with- out changes for the set of j’s which are compatible with symplectic structures in a given deformation class. So

Index π = 2(c1(A) + n) = 2(1 + 2) = 6.

¡ ¢ −1 However, dim RPSL(2, C) = 6 so dim π (j0)/P SL(2, C) = 0 pro- vided j0 is regular. Step 3. In dimension 4 we have the following phenomenon. Two different germs of j-holomorphic curves always intersect with positive index at a common point. This fact is well known for integrable j’s. In the non-integrable case this is a simple linear algebra statement provided the intersection is transversal. However the proof requires a delicate local analysis for non-transversal intersections. As a corol- lary, there is a unique j0-holomorphic curve in A. Since if there were another one, say Σ0, then ([Σ], [Σ0]) = ([Σ], [Σ]) = −1 contradicting 0 −1 that ([Σ], [Σ ]) ≥ 0. We conclude that π (j0)/P SL(2, C) consists of exactly one point. 102 CHAPTER 10. DEFORMATIONS OF SYMPLECTIC FORMS

Step 4. Choose j0 and jt to be regular. We claim that generically no cusp curves appear. Indeed, set A = A1 + ... + Ad, d > 1 and

c1(A) = 1 = c1(A1) + ... + c1(Ad).

Thus at least one of the Chern classes on the right hand side is non- positive. Assume without loss of generality, that c1(A1) ≤ 0 and that A1 is represented by a js-holomorphic curve for some s ∈ [0, 1]. Denote by π the projection in the class A . We see that π−1(γ)/P SL(2, C) A1 1 A1 is non-empty. On the other hand,

dim π−1(γ)/P SL(2, C) = 2(c (A ) + 2) − 6 + 1 ≤ −1, A1 1 1 which is impossible. Thus generically no bubbling off happens (this reflects the fact that codim J 0 ≥ 2). Hence A is represented by a j1-holomorphic curve and therefore (see 10.2.E) ([ω1], [A]) > 0. 2 Chapter 11

An application to ergodic theory

In the present chapter we discuss an asymptotic geometric invariant associated to the fundamental group of Ham(M, Ω) and describe an application to classical ergodic theory (see [P9]).

11.1 Hamiltonian loops as dynamical ob- jects

Let (M, Ω) be a closed symplectic manifold. Given an irrational num- ber α and a smooth loop h : S1 → Ham(M, Ω), one can define a skew 1 1 product map Th,α : M × S → M × S by Th,α(y, t) = (h(t)y, t + α). Our purpose is to relate the geometry and topology of Hamiltonian loops to the dynamics of the associated skew products. 1 The definition above is a particular case of a much more general notion of a skew product (see [CFS, p. 231]) which has been intensively studied during several decades. There are at least two important rea- sons for the interest in this notion. Firstly, it serves as a foundation for mathematical models of random dynamics (see [Ki] for a survey). Secondly, it provides non-trivial examples of systems with interesting

1In this chapter we work with free loops and do not assume that h(0) = 1l.

103 104 CHAPTER 11. ERGODIC THEORY

dynamical properties. The dynamical property we consider is the strict ergodicity. Recall that a T of a compact X is called strictly ergodic if it has precisely one invariant Borel probability mea- sure, say m , which in addition is positive on non-empty open subsets. Strictly ergodic are ergodic, and have a number of additional remarkable features. We mention one of them which plays a crucial role below. Namely, given such a T and an arbitrary F on X, the time averages 1 ΣN−1F (T ix) converge uniformly R N i=0 to the space average X F dm, and in particular converge for all x ∈ X. Note that for general ergodic transformations such a convergence takes place only for almost all x. The contrast between ”all” and ”almost all” becomes especially transparent when one notices that there are purely topological obstructions to the strict ergodicity. For instance, the 2-sphere admits no strictly ergodic homeomorphisms. Indeed the Lefschetz theorem implies that every homeomorphism of S2 has ei- ther a fixed point, or a periodic orbit of period 2 and we see that the invariant measure which is concentrated on such an orbit contradicts the definition of strict ergodicity. In 11.2 below we describe a more sophisticated obstruction to the strict ergodicity which comes from symplectic topology. We say that a loop h : S1 → Ham(M, Ω) is strictly ergodic if for 2 some α the corresponding skew product map Th,α is strictly ergodic. With this language our central question can be formulated as follows. Question. Which homotopy classes S1 → Ham(M, Ω) can be repre- sented by strictly ergodic loops? Here is an example where one gets a complete answer to this question. Let M∗ be the blow up of the complex projective plane 2 CP at one point. Choose a K¨ahlersymplectic structure Ω∗ on M∗ 1 which integrates to 1 over a general line and to 3 over the excep- tional divisor. In view of 9.2.E above one can equivalently think that M∗ = P(T ⊕ C), where T and C are the tautological and the trivial

2 1 Note that each Th,α preserves the canonical measure on M ×S induced by the symplectic form. Thus in our setting the strict ergodicity means that this measure is (up to a factor) the unique invariant measure. 11.2. THE ASYMPTOTIC LENGTH SPECTRUM 105

holomorphic line bundles over CP 1 respectively. With this language the exceptional divisor is our old friend Σ from 10.3, while the gen- eral line is homologous to the sum of Σ with the fiber. The periods of the symplectic form are chosen in such a way that its cohomology class is a multiple of the first Chern class of M (see Definition 11.3.A below). One can easily see that (M∗, Ω∗) admits an effective Hamil- tonian action of the U(2), in other words there exists a monomorphism i : U(2) → Ham(M∗, Ω∗). The fundamental group of U(2) equals Z. It was proved recently by Abreu and McDuff [AM] that the inclusion π1(U(2)) → π1(Ham(M∗, Ω∗)) is an isomorphism, and thus π1(Ham(M∗, Ω∗)) = Z. As far as I know this is the simplest example of a symplectic manifold with π1(Ham) = Z.

Theorem 11.1.A. The trivial class 0 ∈ π1(Ham(M∗, Ω∗)) is the only one which can be represented by a strictly ergodic loop.

The proof of this theorem consists of two parts. First of all, one has to establish existence of contractible strictly ergodic loops. This can be done by purely ergodic methods in a fairly general situation. We refer the reader to [P9] for an extensive discussion and proofs. Secondly, one has to prove that every class γ 6= 0 cannot be represented by a strictly ergodic loop. It turns out that the obstruction comes from the geometry of Ham(M, Ω). We are going to discuss this in more details.

11.2 The asymptotic length spectrum

Define the asymptotic norm of an element γ ∈ π1(Ham(M, Ω)) by ν(kγ) ν∞(γ) = lim , k→+∞ k where ν is the norm introduced in 7.3. This notion is similar to asymp- totic growth µ defined in 8.2 above. The limit exists since the sequence ν(kγ) is subadditive.

Theorem 11.2.A. Let γ ∈ π1(Ham(M, Ω)) be a class represented by a smooth strictly ergodic loop. Then the asymptotic norm ν∞(γ) vanishes. 106 CHAPTER 11. ERGODIC THEORY

Proof: The proof is based on an asymptotic curve shortening procedure in the spirit of 8.3 above. Let h : S1 → Ham(M, Ω) be a smooth loop of Hamiltonian diffeomorphisms which defines a strictly ergodic skew product T (y, t) = (h(t)y, t+α). Let γ be the corresponding element in π1(Ham(M, Ω)). Denote by H(x, t) the normalized Hamiltonian func- −1 −1 tion generating the loop h(t) . Set hk(t) = h(t + kα) and set

fN (t) = h0(t) ◦ ... ◦ hN−1(t).

It follows from 1.4.D that the loop fN is generated by the normalized Hamiltonian function

−1 FN (y, t) = H(y, t) + H(h0(t) y, t + α) + ...

−1 −1 +H(hN−2(t) ◦ ... ◦ h0(t) y, t + (N − 1)α). This expression can be rewritten as follows:

NX−1 k FN (y, t) = H ◦ T (y, t). k=0

Since T is strictly ergodic and the function FN has zero mean we conclude that Z 1 1 max FN (y, t) − min FN (y, t) dt → 0, N 0 y∈M y∈M when N → ∞. But the expression on the left hand side is exactly 1 N length{fN (t)}. Note now that the loop {fN (t)} represents the el- ement −Nγ. Since ν(Nγ) = ν(−Nγ) we get that ν(Nγ)/N tends to zero when N → ∞. This proves that the asymptotic norm of γ vanishes. 2

I do not know the precise value of ν∞(γ) in any example where this quantity is strictly positive (for instance, for the blow up of CP 2 in 11.1 above). The difficulty is the following. In all known examples where Hofer’s norm ν(γ) can be computed precisely there exists a closed loop h(t) which minimizes the length in its homotopy class 11.2. THE ASYMPTOTIC LENGTH SPECTRUM 107

(that is a minimal closed geodesic). It turns out however that every non-constant minimal closed geodesic loses minimality after a suitable number of iterations. In other words the loop h(Nt) can be shortened provided N is large enough. The proof of this statement is based on the following generalization of the shortening procedure described above. Let H(y, t) be the normalized Hamiltonian of h(t)−1. Assume without loss of generality that h(0) = 1l, and that H(y, 0) does not vanish identically. Denote by Γ the set of all points of M where the function |H(y, 0)| attains the maximal value. Since M \ Γ is a non-empty open subset, and the group of Hamiltonian diffeomorphisms acts transitively on M, one can choose a sequence

φ0 = 1l, φ1, ..., φN−1 ∈ Ham(M, Ω)

such that Γ ∩ φ1(Γ) ∩ ... ∩ φN−1(Γ) = ∅. −1 −1 −1 −1 −1 Consider the loop fN (t) = h(t) ◦ φ1h(t) φ1 ◦ ... ◦ φN−1h(t) φN−1. We claim that it is shorter than the loop h(Nt). Indeed, note that its Hamiltonian FN at time t = 0 can be written as follows: NX−1 −1 FN (y, 0) = H(φi y, 0). i=0 Set a(t) = max FN (y, t) − min FN (y, t) y∈M y∈M and b(t) = N(max H(y, t) − min H(y, t)). y∈M y∈M

Our choice of the sequence {φi} implies that a(0) < b(0). Since a(t) ≤ R 1 R 1 b(t) for all t, we get that 0 a(t)dt < 0 b(t)dt, and this proves the claim. We conclude that if a non-zero class γ ∈ π1(Ham(M, Ω)) is represented by a minimal closed geodesic then ν∞(γ) is strictly less than ν(γ). It would be interesting to investigate further restrictions on the homotopy classes of smooth strictly ergodic loops in the group of Hamiltonian diffeomorphisms. 108 CHAPTER 11. ERGODIC THEORY 11.3 Geometry via algebra

Let us return to Theorem 11.1.A. In this section we very briefly sketch the proof of the fact that non-contractible loops on Ham(M∗, Ω∗) can- not be strictly ergodic. Let (M 2n, Ω) be a closed symplectic manifold. Define a mapping

I : π1(Ham(M, Ω)) → R as follows. Take an element γ ∈ π1(Ham(M, Ω)), and consider the associated symplectic fibration P (γ). Denote by u the first Chern class of the vertical tangent bundle to P (γ). The fiber of this bundle at a point of P (γ) is the (symplectic) vector space tangent to the fiber through this point. As before, we write c for the coupling class. Define the “characteristic number” Z I(γ) = cn ∪ u. P (γ)

It is easy to see that I : π1(Ham(M, Ω)) → R is a homomorphism ([P6],[LMP2]). Definition 11.3.A. A symplectic manifold (M, Ω) is called monotone if [Ω] is a positive multiple of c1(TM). Theorem 11.3.B. ([P6]) Let (M, Ω) be a closed monotone symplec- tic manifold. Then there exists a positive constant C > 0 such that ν(γ) ≥ C|I(γ)| for all γ ∈ π1(Ham(M, Ω)). In other words, the homomorphism I calibrates Hofer’s norm on the fundamental group. The proof of this theorem is based on the theory described in two previous chapters combined with results from [Se]. Recently Seidel obtained a generalization of this inequality to non-monotone symplectic manifolds. Since I is a homomorphism, the estimate in 11.3.B extends to the asymptotic Hofer’s norm:

ν∞(γ) ≥ C|I(γ)|. The great advantage of the homomorphism I is that it can be relatively easily calculated in examples. For instance, one can show 11.3. GEOMETRY VIA ALGEBRA 109

that I(γ) 6= 0 where γ is the generator of π1(Ham(M∗, Ω∗)) = Z. Thus 11.3.B implies that the asymptotic norm of every non-trivial element of π1(Ham(M∗, Ω∗)) is strictly positive. We conclude from 11.2.A that such an element cannot be represented by a strictly ergodic loop. 110 CHAPTER 11. ERGODIC THEORY Chapter 12

Elements of variational theory of geodesics

We have already seen a number of results concerning length minimiz- ing geodesics on the group of Hamiltonian diffeomorphisms. In the present chapter we discuss a different viewpoint on geodesics based on calculus of variations. Given a smooth path of Hamiltonian dif- feomorphisms, is it possible to shorten it by a small variation with fixed endpoints? This question is motivated by the classical theory of geodesics on Riemannian manifolds. Interestingly enough, in Hofer’s geometry at least under certain non-degeneracy assumptions, one can give a rather precise answer with a transparent dynamical meaning [U] .

12.1 What are geodesics?

The intuition coming from suggests that the geodesics should be defined as critical points of the length functional. Let us try to formalize this definition.

Let {ft}, t ∈ [a; b] be a smooth path in Ham(M, Ω). A variation of {ft} is a smooth family of paths {ft,ε} with t ∈ [a; b] and ε ∈ (−ε0, ε0) satisfying fa,ε = fa, fb,ε = fb and ft,0 = ft for all t and ε. We always assume that ∪supp ft,ε is compact. Given such a variation, consider t,ε

111 112 CHAPTER 12. GEODESICS

the length of the path {ft,ε} as the function of ε:

Z b Z b `(ε) = ||F (·, t, ε)||dt = max F (x, t, ε) − min F (x, t, ε)dt, a a x x

where F (x, t, ε) is the Hamiltonian generating the path {ft,ε} for given ε.

Tentative Definition 12.1.A. A path {ft} is a geodesic if • it has constant speed, that is ||F (·, t)|| does not depend on t;

• for every smooth variation of {ft} the length function `(ε) has a critical point at ε = 0.

Here we immediately face a difficulty - even for smooth variations the function `(ε) need not be smooth! Thus the notion of a critical point should be specified. Our first task is to figure out the structure of the length functions `(ε) associated to variations of the given path. Proposition 12.1.B. Every length function `(ε) is convex at 0 up to second order terms. Namely, there exists a convex function u(ε) and numbers δ > 0, C > 0 such that

|`(ε) − u(ε)| ≤ Cε2 for all ε ∈ (−δ, δ). Note that whatever the definition of the critical point is, it should not depend on the second order terms. The only natural candidate for the critical point of a convex function is its minimum. Thus we arrive at the following notion. Definition 12.1.C. The point ε = 0 is a critical point for the length function `(ε) if and only if it is a point of minimum for a convex function u(ε) which satisfies inequality 12.1.B. Exercise 12.1.D.

• The definition above is correct, that is it does not depend on the particular choice of a convex function u which satisfies 12.1.B. 12.1. WHAT ARE GEODESICS? 113

• If `(ε) is smooth, Definition 12.1.C coincides with the usual one. • If ` attains its local minimum or maximum at 0 then 0 is a critical point of ` in the sense of 12.1.C. • Conclude from 12.1.B that the germ of the function 1 − |ε| at 0 cannot arise as the length function of any smooth variation.

Now we are ready to answer the question posed in the title of this section : a path {ft}, t ∈ [a; b] is called a geodesic if it satisfies 12.1.A and 12.1.C. It follows from 12.2.A below that the restriction of a geodesic defined on [a; b] to any subsegment is again a geodesic. Thus the notion of geodesic extends in the obvious way to paths which are defined on arbitrary time intervals. Let I ⊂ R be an interval, and let f : I → Ham(M, Ω) be a smooth path. If the restriction of f to any segment [a; b] ⊂ I is a geodesic the path f itself is called a geodesic. We will focus on geodesics {ft} which are defined on the unit interval [0; 1]. Moreover since Hofer’s metric is biinvariant we can always shift a geodesic and assume that f0 = 1l. Let us turn to the proof of 12.1.B. It is convenient to use the linearization idea introduced in Chapter 5. Consider the space F0 of all normalized Hamiltonians M × [0; 1] → R endowed with the norm Z 1 |||F |||0 = max F (x, t) − min F (x, t)dt. 0 x x

Denote by H0 ⊂ F0 the subset consisting of all Hamiltonians which generate a loop {ht} of Hamiltonian diffeomorphisms: h0 = h1 = 1l. Let us emphasize that the functions from F0 and H0 are not assumed to be periodic in time, in the contrast with the setting of Chapter 5. Denote by V the set of all smooth families H(x, t, ε) of functions from H0 with H(x, t, 0) ≡ 0.

Proposition 12.1.E. Let {ft} be a smooth path of Hamiltonian dif- feomorphisms with t ∈ [0; 1] and f0 = 1l. The set of length functions `(ε) associated with the variations of {ft} consists of all functions of the form |||F − H(ε)|||0, 114 CHAPTER 12. GEODESICS where H ∈ V.

Proposition 12.1.B is an immediate consequence of this formula. 0 Indeed, set u(ε) = |||F − εH (0)|||0. Clearly u is convex and coincides with ` up to the second order terms.

Proof of 12.1.E: Every variation of the path ft can be written in −1 the form ft,ε = ht,ε ◦ ft, where ht,ε is a smooth family of loops with ht,0 = 1l. The Hamiltonian H(x, t, ε) of {ht,ε} belongs to V. Further,

F (x, t, ε) = −H(ht,εx, t, ε) + F (ht,εx, t),

thus

`(ε) = |||F − H(ε)|||0. 2

Solve the following exercises using 12.1.E. Exercise 12.1.F. Assume that for every t the function F (x, t) has a unique point of maximum and minimum respectively, and moreover these points are non-degenerate in the sense of . Then for every variation of {ft} the length function `(ε) is smooth. Hint: Use the implicit function theorem. Exercise 12.1.G. Assume that the maximum set of F has non-empty interior. Construct a variation such that the corresponding length function is not smooth at 0.

12.2 Description of geodesics

We say that a function F ∈ F0 has fixed extrema if there exist two points x−, x+ ∈ M such that F (x−, t) = minx F (x, t) and F (x+, t) = maxx F (x, t) for all t ∈ [0; 1], and moreover the function F (x+, t) − F (x−, t) does not depend on t. The crucial point here is that the extremal points x− and x+ of F (·, t) are time-independent. 12.2. DESCRIPTION OF GEODESICS 115

Theorem 12.2.A. A path {ft} is a geodesic if and only if the corre- sponding Hamiltonian function F ∈ F0 has fixed extrema.

In particular, every autonomous Hamiltonian flow is a geodesic. Hamiltonian functions with fixed extrema were introduced in [BP1] where they were called quasi-autonomous. Theorem 12.2.A above is essentially due to [LM2] (though the definition of a geodesic is given there in a slightly different way). In order to prove Theorem 12.2.A we have to investigate in more details the structure of variations. Denote by V1 the tangent space to H0 at 0: ∂H ¯ V = { (0)¯H ∈ V}. 1 ∂ε

Proposition 12.2.B. The space V1 consists of all functions G ∈ F0 which satisfy Z 1 G(x, t)dt = 0 0 for all x ∈ M.

The proof of the proposition is absolutely analogous to 5.2.D and 6.1.C above where the case of time-periodic variations was settled.

Proof of Theorem 12.2.A. Let F ∈ F0 be a function such that ||F (·, t)|| does not depend on t. For a family H(ε) from V put G = 0 H (0) ∈ V1. Define the functions `(ε) = |||F − H(ε)|||0 and v(ε) = |||F −εG|||0. In view of 12.1.E the Hamiltonian F generates a geodesic if and only if `(ε) has a critical point in the sense of 12.1.C at ε = 0 for every H ∈ V. Since `(ε) and v(ε) coincide up to the second order terms, and v(ε) is convex, this is equivalent to the fact that v(ε) has a point of minimum at ε = 0 for every G ∈ V1. Therefore in order to prove Theorem 12.2.A it suffices to verify that the following conditions are equivalent: (i) F has fixed extrema;

(ii) |||F − εG|||0 ≥ |||F |||0 for all G ∈ V1. 116 CHAPTER 12. GEODESICS

Assume that (i) holds. Set u = F − εG for G ∈ V1. Then Propo- sition 12.2.B implies that

Z 1 |||u|||0 ≥ u(x+, t) − u(x−, t)dt = |||F |||0, 0 and hence we get (ii). Assume now that (ii) holds. Take

Z 1 G(x, t) = F (x, t) − F (x, t)dt. 0 The inequality (ii) yields

Z 1 Z 1 Z 1 max F (x, t)dt−min F (x, t)dt ≥ max F (x, t)−min F (x, t)dt, x 0 x 0 0 x x which is possible only if F has fixed extrema. 2

12.3 Stability and conjugate points

A geodesic is called stable if for every variation the length function `(ε) attains its minimal value at 0. In other words, stable geodesics cannot be shortened by small variations with fixed end points. The problem of characterization of stable geodesics in full generality re- mains open. 1 Below we present a solution for a special class of non- degenerate geodesics. By definition, a geodesic is non-degenerate if for every t the corresponding Hamiltonian function has a unique point of maximum and minimum respectively, and moreover these points are non-degenerate in the sense of Morse theory. For instance, every au- tonomous path with unique non-degenerate maximum and minimum

1I believe that it can be solved by existing methods of non-smooth analysis. Note also that the first statement of Theorem 12.3.A below which provides a sufficient condition for stability holds true for all geodesics, see [LM3]. The proof involves theory of pseudo-holomorphic curves. 12.4. THE SECOND VARIATION FORMULA 117

is a non-degenerate geodesic. Recall that for every variation of a non- degenerate geodesic the length function `(ε) is smooth (see 12.1.F). The theory developed in this and the next two sections is due to Ustilovsky [U] (see also [LM2]). Let {ft}, t ∈ [0; 1], f0 = 1lbe a non-degenerate geodesic generated by a Hamiltonian F ∈ F0. Denote by x− and x+ the time-independent points of minimum and maximum of F (·, t) respectively. Note that x± are fixed points of {ft} in this case. Consider the linearized flows ft∗ on

Tx+ M and Tx− M. We say that such a flow has a non-trivial T -periodic orbit with T 6= 0 if fT ∗ξ = ξ for some tangent vector ξ 6= 0. Theorem 12.3.A. ([U]) • Assume that the linearized flows have no non-trivial T -periodic orbits with T ∈ (0, 1]. Then {ft} is stable.

• Assume that {ft} is stable. Then the linearized flows have no non-trivial periodic orbits with T ∈ (0, 1). This result can be interpreted as a description of conjugate points along geodesics of Hofer’s metric. Namely, conjugate points correspond to non-trivial closed orbits of the linearized flow at x±. Before the conjugate point the geodesic is stable and after the conjugate point it loses its stability (that is it can be shortened by a small variation). We refer to [U] for further information. The proof of 12.3.A is presented in 12.5 below. It is based on the second variation formula which we are going do describe in the next section. Exercise 12.3.B. Conclude from 12.3.A that a sufficiently short seg- ment of a non-degenerate geodesic is stable.

12.4 The second variation formula

Let {ft} be a non-degenerate geodesic generated by the Hamiltonian F (x, t) with maximum/minimum points x±. Denote by C±(t) the op- erator of the linearized equation at x± i.e. d f (x ) = C (t)f (x ). dt t∗ ± ± t∗ ± 118 CHAPTER 12. GEODESICS

Consider the spaces ¯ ¯ V± = {smooth maps v : [0, 1] → Tx± M ¯ v(0) = v(1) = 0}.

Given a variation {ft,ε} of {ft}, define the elements v± ∈ V± by ¯ d ¯ v± = ¯ ft,εx±. dε ε=0 It will be convenient to consider the maximum and the minimum part of the length function associated to the variation separately. Set

Z 1 `+(ε) = max F (x, t, ε)dt, 0 x and Z 1 `−(ε) = min F (x, t, ε)dt. 0 x

Clearly `(ε) = `+(ε) − `−(ε). Theorem 12.4.A. ([U])

d2` ± (0) = Q (v ) dε2 ± ± and hence d2` (0) = Q (v ) − Q (v ), dε2 + + − − where Z 1 ³ ´ −1 Q±(v) = − Ω(C± v,˙ v˙) + Ω(v, ˙ v) dt. 0 Example (Isoperimetric inequality). Consider the standard sym- plectic plane R2(p, q) endowed with the symplectic form ω = dp ∧ dq. Let v : [0, 1] → R2 be a smooth curve with v(0) = v(1) = 0. Recall the following notions from Euclidean geometry:

Z 1 length (v) = |v˙|dt, 0 12.4. THE SECOND VARIATION FORMULA 119

Z 1 energy (v) = |v˙|2dt, 0 Z 1 1 area (v) = h (pdq − qdp), v˙idt 2 0Z 1 1 = pq˙ − qp˙ dt 2 Z0 1 1 = ω(v, v˙)dt, 2 0 where v(t) = (p(t), q(t)). The isoperimetric inequality reads 4π area (v) ≤ ( length (v))2 ≤ energy (v). Going back to our symplectic setting, we assume that the Hamiltonian 2 2 near x− is given by F (x) = πλ(p + q ) with λ > 0. So we get the Hamiltonian system ½ p˙ = −2πλq q˙ = 2πλp Setting z = p+iq we obtain the linear equationz ˙ = 2πλiz. In this case −1 i 2 C−(t) = 2πλi and C− (t) = − 2πλ and, recalling that ω(ξ, iξ) = |ξ| , we get Z 1 ¡ i ¢ Q (v) = − ω(− v,˙ v˙) + ω(v, ˙ v) dt − 2πλ 0 Z Z 1 1 1 = − |v˙|2dt − ω(v, ˙ v)dt 2πλ 0 0 1 = − energy (v) + 2 area (v). 2πλ Equationz ˙ = 2πλiz has the solution z(t) = e2πλitz(0) for t ∈ [0, 1] and thus it has no orbits of period 1 for λ < 1. Combining Theorems 12.3.A and 12.4.A we get that Q−(v) ≤ 0 for all plane curves v with v(0) = v(1) = 0. Therefore 4πλ area (v) ≤ energy (v) for all λ < 1. Letting λ go to 1 we get the isoperimetric inequality. For the proof of 12.4.A we need the following auxiliary statement. 120 CHAPTER 12. GEODESICS

Lemma 12.4.B. Let {ft,ε} be a variation of the path {ft}, and let F (x, t, ε) be the corresponding Hamiltonian function. Then Z 1 ∂F (ft,εx, t, ε)dt = 0. 0 ∂ε

Proof: 1) The formula above is valid for any variation of the constant loop, that is when ft = 1l for all t. The proof is analogous to the proof of 6.1.C above.

2) Consider now the general case. Write ft,ε = ft◦ht,ε, where ht,ε is a variation of the constant loop. Denote by H(x, t, ε) the Hamiltonian −1 of {ht,ε}. Then F (x, t, ε) = F (x, t) + H(ft x, t, ε), and hence

∂F ∂H (x, t, ε) = (f −1x, t, ε). ∂ε ∂ε t

Therefore ∂F ∂H (f x, t, ε) = (h x, t, ε). ∂ε t,ε ∂ε t,ε The required statement follows now from step 1. 2

Proof of 12.4.A: We will calculate the second derivative of `+(ε). The function `−(ε) is treated analogously. We work in the standard sym- plectic coordinates x = (p, q) near x+ and write ξ · η for the Euclidean scalar product. Let i denote the complex structure (p, q) 7→ (−q, p). In this terminology Ω(ξ, iη) = ξ · η and the Hamiltonian system reads

d ∂F f x = i (f x, t). dt t ∂x t

Thus ∂2F C (t) = i (x , t). + ∂x2 + To simplify the formulae later on, we will introduce the following no- 12.4. THE SECOND VARIATION FORMULA 121

tation: ∂2F a = (x , t, 0) ∂x∂ε + ∂x b = + (t, 0) ∂ε ∂2F c = (x , t, 0) ∂ε2 + ∂2F K = (x , t). ∂x2 + The implicit function theorem insures that F (·, t, ε) has the unique point of maximum x+(t, ε) which depends smoothly on t and ε. Since

Z 1 `+(ε) = F (x+(t, ε), t, ε)dt 0 we get that

Z 1 d`+ ∂F ∂x+ ∂F (ε) = (x+(t, ε), t, ε) (t, ε) + (x+(t, ε), t, ε)dt. dε 0 ∂x ∂ε ∂ε

Since x+(t, ε) is a critical point of F (·, t, ε) this expression simplifies to Z 1 d`+ ∂F (ε) = (x+(t, ε), t, ε)dt. dε 0 ∂ε Differentiating again with respect to ε and setting ε = 0 we get

2 Z 1 d `+ 2 (0) = (a · b + c)dt. (12.4.C) dε 0

∂F Differentiation of the equation ∂x (x+(t, ε), t, ε) = 0 with respect to ε at ε = 0 yields Kb + a = 0. The non-degeneracy condition guarantees that we can invert K and thus b = −K−1a. (12.4.D) 122 CHAPTER 12. GEODESICS

In view of Lemma 12.4.B Z 1 ∂F (ft,εx, t, ε)dt = 0. 0 ∂ε

Differentiating this with respect to ε at ε = 0 and x = x+ we get Z 1 (a · v+ + c)dt = 0. (12.4.E) 0 Consider the Hamilton equation d ∂F f x = i (f x, t, ε). dt t,ε ∂x t,ε

Differentiate it with respect to ε at ε = 0 and x = x+. We obtain

v˙+ = iKv+ + ia or equivalently a = −iv˙+ − K v+. ( 12.4.F) Taking 12.4.D and 12.4.F together, we have −1 −1 b = −K (−iv˙+ − Kv+) = K iv˙+ + v+. In view of 12.4.C this implies 2 Z 1 d `+ 2 (0) = (a · b + c)dt dε 0 Z 1 −1 = (a · K iv˙+ + a · v+ + c)dt 0 Z 1 12.4.E −1 = a · K iv˙+dt 0 Z 1 12.4.F −1 = − (iv˙+ + Kv+) · K iv˙+dt 0 Z 1 ∗ −1 = − (K iv˙+ · iv˙+ + v+ · iv˙+)dt 0 Z 1 −1 = − (−C+ v˙+ · iv˙+ + v+ · iv˙+)dt 0 Z 1 −1 = − (Ω(C+ v˙+, v˙+) + Ω(v ˙+, v+))dt. 0 12.5. ANALYSIS OF THE SECOND VARIATION FORMULA 123

The equality (∗) follows from the fact that K is symmetric. Further- −1 −1 more we have used that C+ = −K i and Ω(ξ, iη) = ξ · η (or equiv- alently −Ω(ξ, η) = ξ · iη). 2

12.5 Analysis of the second variation for- mula

Proposition 12.5.A. For all a ∈ V+ and for all b ∈ V− there exists a variation {ft,ε} of {ft} such that v+ = a and v− = b.

Proof: Take a function G ∈ F0 such that

Z 1 G(x, t)dt = 0 0

for all x ∈ M (that is G ∈ V1 in the notation of 12.2). This choice will be specified later on. Define the variation {ht,ε} of the constant loop ht,0 = 1l as follows: ht,ε is the time-ε-map of the Hamiltonian R t 0 G(x, s)ds. Consider a variation ft,ε = ftht,ε of {ft}. We have ¯ d ¯ v+(t) = ¯ ft,εx+ dε ε=0 ¯ d ¯ = ¯ ftht,εx+ dε ε=0 ¯ d ¯ = ft∗ ¯ ht,εx+. dε ε=0

Our explicit construction of {ht,ε} yields

¯ Z t d ¯ ¯ ht,εx+ = sgrad G(x+, s)ds. dε ε=0 0 Set d c(t) = f −1a(t). dt t∗

Choose now G ∈ V1 such that in the standard symplectic coordinates near x+ G(x, t) = Ω(x − x+, c(t)). 124 CHAPTER 12. GEODESICS

Since a(0) = a(1) = 0 by definition of V+ this condition is compatible with the fact that G ∈ V1. The explicit expression for G yields d sgrad G(x , t) = c(t) = f −1a(t). + dt t∗ Integrating this equality and taking into account that a(0) = 0 we get that Z t −1 sgrad G(x+, s)ds = ft∗ a(t). 0

This implies that v+(t) = a(t). The same argument works for x− . Since x+ and x− are disjoint, the variations we constructed do not interfere. This completes the proof. 2

Proof of Theorem 12.3.A: In view of 12.4.A and 12.5.A we come to the conclusion that {ft} is stable if and only if both functionals Q+ and −Q− are non-negative. Classical calculus of variations provides us with precise conditions when this happens. In what follows we will focus on Q+. The same arguments work for Q−. −1 Consider the Lagrangian L(v, ˙ v) = −Ω(C+ v,˙ v˙) − Ω(v, ˙ v). Let us investigate the critical points of the functional

Z 1 L(v ˙(t), v(t))dt 0 on V+. First of all, we claim that L satisfies the Legendre condition, ∂2L −1 that is ∂v˙ 2 is positive definite. Indeed, L(v, ˙ v) = −iC+ v˙ · v˙ − iv˙ · v ∂2L −1 ∂2F −1 and therefore ∂v˙ 2 = −2iC+ . Since C+ = i ∂x2 (x+) we get −iC+ = ³ ´−1 ∂2F i ∂x2 (x+) i and so

D ³∂2F ´−1 E D³∂2F ´−1 E i (x ) iξ, ξ = − (x ) iξ, iξ . ∂x2 + ∂x2 +

The point x+ is a non-degenerate maximum of F . Thus the ³ ´−1 ∂2F ∂2F ∂x2 (x+) is negative definite and so is ∂x2 (x+) . We deduce that ∂2L ∂v˙ 2 is positive definite. The claim follows. 12.6. LENGTH MINIMIZING GEODESICS 125

Since L is quadratic with respect to v, the constant path v0(t) ≡ 0 is an extremal. For quadratic functionals satisfying the Legendre condition, classical calculus of variations tells us the following. If v0 is the minimum then for all T ∈ [0, 1) the Euler-Lagrange equation (which coincides with the Jacobi equation) has no non-trivial solutions with v(0) = v(T ) = 0. Vice versa, if there are no such solutions for T ∈ [0, 1] then v0 is the minimum. It remains to relate the Euler-Lagrange

equation to periodic orbits of the linearized flow ft∗ on Tx+ M. The Euler-Lagrange equation reads d ∂L ∂L = ⇐⇒ dt ∂v˙ ∂v d (−2iC−1v˙ + iv) = −iv˙ ⇐⇒ dt + d d (−2iC−1v˙) = (−2iv). dt + dt −1 If v is a solution then C+ v˙ = v + const. Setting w = v + const we have ½ −1 C+ w˙ = w w(0) = w(T ). But this means precisely that w is a T -periodic orbit of the linearized flow. This proves Theorem 12.3.A. 2

12.6 Length minimizing geodesics

Let I ⊂ R be an interval. Consider a geodesic {ft}, t ∈ I such that all its tangent vectors have unit speed: d || f || = 1 dt t for all t ∈ I. Such a geodesic is called locally strictly minimal (cf. 8.2 above) if for every t ∈ I there exists a neighbourhood U of t in I such that ρ(fa, fb) = |a − b| 126 CHAPTER 12. GEODESICS for all a, b ∈ U. Every geodesic on a is locally strictly min- imal. This is essentially due to the fact that the exponential map of the Levi-Civita connection is a diffeomorphism on a neighbourhood of zero. There is no analogue of such a fact in Hofer’s geometry. Still various examples serve as a motivation for the following conjecture. Conjecture 12.6.A. Every geodesic of Hofer’s metric is locally strictly minimal. This was proved first for the standard symplectic R2n in [BP1]. Later this conjecture was confirmed in [LM2] for some other symplec- tic manifolds including cotangent bundles, closed oriented surfaces and CP 2. It is useful to relax the notion of local strict minimality in the following way. We say that a geodesic {ft}, t ∈ I is locally minimal if for every t ∈ I there exists a neighbourhood U of t in I such that for every a, b ∈ U with a < b the geodesic segment {ft}, t ∈ [a, b] minimizes the length in the homotopy class of paths with fixed end points. Conjecture 12.6.B. Every geodesic of Hofer’s metric is locally min- imal. In finite-dimensional Riemannian geometry the local minimality yields the local strict minimality. In Hofer’s geometry this is true un- der the following additional assumption. Denote by S ⊂ [0; +∞) the length spectrum of Ham(M, Ω) (see 7.3 above). Set a = inf(S \{0}). As usual the infimum of the empty set is assumed to be +∞.

Proposition 12.6.C. ([LM2]) Assume that a > 0. Then every locally minimal geodesic is in fact locally strictly minimal.

Proof: Let {ft} be a locally minimal geodesic. Assume without loss of generality that its Hamiltonian has unit norm at each time moment t. Thus there exists ε ∈ (0; a/2) with the following property. For every path α on Ham(M, Ω) which joins 1l with fε and which is homotopic to {ft}, t ∈ [0; ε] one has length α ≥ ε. It suffices to prove that ρ(1l, fε) = ε. Assume on the contrary that there exists a path β which 12.6. LENGTH MINIMIZING GEODESICS 127

joins 1l with fε such that length β < ε. Write γ for the path {ft}, t ∈ [0; ε]. Consider a loop formed by γ and β. Its length is strictly less than a. Hence, by the definition of a, this loop can be homotoped to an arbitrarily short loop. But this implies that the path γ can be homotoped to a path γ0 whose length is arbitrarily close to the length of β. We get a contradiction with the inequalities length γ0 ≥ ε and ε > length β. 2

In particular, if for a manifold (M, Ω) the quantity a is strictly positive then 12.6.B implies 12.6.A. The condition a > 0 is hard to verify. It holds for Liouville manifolds (see 7.3.B above), in the case 2 when π1(Ham(M, Ω)) is finite (for instance, for surfaces or CP ), and in a few other examples in dimension 4. No example with a = 0 is known. Conjecture 12.6.B is confirmed in [LM2] for an interesting class of symplectic manifolds. Let us describe this class in more details. Assume that M is semi-monotone (cf. 11.3.A), that is there exists a constant k ≥ 0 such that c1(A) = k([Ω],A) for every spherical homology class A ∈ H2(M, Z). Further, when M is open, assume in addition that M is ”geometrically bounded” at infinity. For instance any cotangent bundle T ∗N endowed with the standard symplectic structure is geometrically bounded. For the precise definition see [AL]. Then 12.6.B holds. Let us turn to a question which we started to discuss in 8.2 above. Given a geodesic, what can one say about the length of the time interval on which it is minimal? There exists a beautiful approach to this question based on the careful study of closed orbits of the corresponding Hamiltonian flow. It becomes especially transparent in the case of autonomous geodesics, that is one-parameter subgroups of Ham(M, Ω). Let {ft} be a one-parameter subgroup of Ham(M, Ω). The standard ODE technique yields that there exists a strictly positive number τ > 0 with the following property (see [HZ, Section 5.7]). Every closed orbit of the flow {ft} on the time interval [0; τ] is a fixed point of the flow. Conjecture 12.6.D. Every one-parameter subgroup is minimal on the interval [0; τ]. 128 CHAPTER 12. GEODESICS

This phenomenon was discovered by Hofer in [H2] for the case of R2n. It could be considered as the global version of the conjugate points criterion 12.3.A above. 2 For generalizations to non-autonomous flows as well as to other symplectic manifolds we refer to [BP1], [Si1], [LM2], [Sch3] and [Sli]. Conjecture 12.6.D provides an interesting existence mechanism for non-trivial closed orbits of autonomous Hamiltonian flows. Indeed, assume that we know a priori that the autonomous flow {ft} is not minimal on some time interval. Such an information can come for instance from conjugate points, or from shortening proce- dures described in 8.2, 8.3 above. Then 12.6.D guarantees the exis- tence of non-trivial closed orbits on that interval. We refer to [LM2] and [P8] for further discussion. The study of length minimizing properties of geodesics led to the understanding of the following surprising feature of Hofer’s geometry. 12.6.E. The C1-flatness phenomenon, [BP1]. There exists a C1- neighbourhood E of the identity in the group Ham(R2n) and a C2- neighbourhood C of 0 in its Lie algebra A such that (E, ρ) is isometric to (C, || ||). Some generalizations can be found in [LM2] . It is instructive to compare the C1-flatness phenomenon with Sikorav’s theorem 8.2.A which states that every one-parameter subgroup of Ham(R2n) remains a bounded distance of the identity. This can be interpreted as a pos- itive curvature type effect which intuitively contradicts the flatness! It is still unclear how to resolve this paradox. We refer to [BP1],[HZ] for the proof of the C1-flatness phenomenon (see also [P8] for further discussion). There are several different approaches to results on local minimal- ity presented above. All of them are rather complicated. In the next chapter we present one of them which as we believe could be developed further in order to prove Conjecture 12.6.B in full generality. We will discuss the following simplest case only.

Theorem 12.6.F. Let (M, Ω) be a closed symplectic manifold with π2(M) = 0. Then every one-parameter subgroup of Ham(M, Ω) gen-

2The discovery of conjugate points was in fact motivated by Hofer’s result. 12.6. LENGTH MINIMIZING GEODESICS 129 erated by a generic time-independent Hamiltonian on M is locally minimal.

The proof is sketched in 13.4 below. 130 CHAPTER 12. GEODESICS Chapter 13

A visit to Floer homology

In the present chapter we sketch a proof of Theorem 12.6.F which states that every one-parameter subgroup of Ham(M, Ω) generated by a generic Hamiltonian function is locally minimal provided π2(M) = 0. Our approach is based on Floer homology. The exposition is nor com- plete neither 100% rigorous. Its purpose is to sketch a very complicated and still developing machinery rather than to make a systematic in- troduction. We follow closely two papers by M. Schwarz [Sch2, Sch3].

13.1 Near the entrance

Floer homology is one of the most powerful tools of modern symplectic topology. Its creation was motivated by the following question: what are invariants of Hamiltonian diffeomorphisms? Let us describe the plan of the answer. We work on a connected closed symplectic mani- fold (M, Ω) assuming for simplicity that the manifold is aspherical: π2(M) = 0. The role of this assumption will be clear in a few mo- ments. Let us introduce some notations. Denote by Ham(] M, Ω) the universal cover of the group of Hamiltonian diffeomorphisms with the base point 1l. Write LM for the space of contractible loops S1 → M. Write LHam(M, Ω) for the group of contractible loops {ht} of Hamil- tonian diffeomorphisms based at 1l generated by functions from H. To simplify the notations, we will often omit the dependence on M and

131 132 CHAPTER 13. FLOER HOMOLOGY

Ω and write Ham for Ham(M, Ω), etc. The group LHam acts naturally on LM by the operators

Th : z(t) → htz(t).

The starting observation: there exists a natural map Ham] → C∞(LM)/LHam. Let us describe this map. Fix an element φ ∈ Ham.] Denote by F(φ) the set of all Hamiltonians F ∈ F which generate φ. The group LHam acts transitively on F(φ). This action is defined as follows. Take a loop h ∈ LHam and a Hamiltonian F ∈ F(φ). Write {ft} for the Hamiltonian flow of F . Then h(F ) is defined as the −1 normalized Hamiltonian function which generates the flow ht ◦ ft. Formula 1.4.D implies that

h(F )(x, t) = −H(htx, t) + F (htx, t).

Given F ∈ F(φ), define a function AF : LM → R called the symplectic action:

Z 1 Z AF (z) = F (z(t), t)dt − Ω, 0 D where D is a disc in M bounded by the loop z. Since M is aspherical, the definition above does not depend on the choice of D. Exercise 13.1.A. Prove that

−1 ∗ (Th ) AF = Ah(F ) for all h ∈ LHam and F ∈ F(φ). Hint: Use that for every contractible loop {ht} ∈ LHam generated by some H ∈ H the action vanishes identically on orbits: AH ({htx}) = 0 for all x ∈ M. In fact on aspher- ical manifolds this is true even for non-contractible loops {ht}. This difficult result was proved recently by Schwarz [Sch3]. This completes the description of the natural map which sends φ ∈ ∞ Ham] to the equivalence class [AF ] ∈ C (LM)/LHam. A function up to a diffeomorphism is a very rich object. For in- stance, in finite dimensions, a lot of information can be extracted from 13.2. MORSE HOMOLOGY IN FINITE DIMENSIONS 133

its critical points and the topology of the level sets. A powerful tool to get such information is Morse theory. In our situation, we have to deal with an infinite-dimensional manifold — the loop space LM. The fundamental observation due to Floer is that there exists a suit- able version of Morse theory which can be extended to this infinite- dimensional setting. We describe it in the next sections. Morse-Floer theory gives rise to a quite complicated structure associated canoni- cally with Ham.] Our main task is to figure out the place of Hofer’s norm of φ inside this structure. As we will see, it is closely related to values of the action functionals at so-called homologically essential critical points. We start with a digression to the finite dimensional situation. 1 The following remark will presumably help the reader to develop the right intuition. The manifold M can be naturally identified with the subset of LM consisting of constant loops. When the function F ∈ F is time-independent, the restriction of AF to M is simply F . Thus the usual function theory on M ”sits inside” the theory of the action functionals on LM.

13.2 Morse homology in finite dimensions

Let F be a Morse function on a closed connected N-dimensional mani- fold M. We write CritF for the set of critical points of F . Denote by 2 i(x) the Morse coindex of a critical point x, and by CritmF the set of critical points with Morse coindex m. Denote by C(F ) the vector space over Z2 generated by CritF , and by Cm(F ) its subspace generated by critical points of coindex m. Take a generic 3 Riemannian metric r on M and consider the

1See the book [Sch1]. 2 2 This is the number of positive squares in the normal form of dxF . 3Here and below ”generic” should be understood as in the last paragraph of 4.2: generic metric is an element of a residual subset in the space of all metrics, etc. 134 CHAPTER 13. FLOER HOMOLOGY

negative gradient flow

du (s) = −∇ F (u(s)). ds r

Pick two points x−, x+ ∈ CritF . Fact 13.2.A. The space of orbits u(s) of the gradient flow which satisfy u(s) → x− as s → −∞ and u(s) → x+ as s → +∞ is a smooth manifold of dimension i(x−) − i(x+). Note that this space carries a natural free R-action. Indeed, if u(t) is a solution, then u(t + const) is a solution as well. Thus when i(x−) − i(x+) = 1 the quotient space is a 0-dimensional manifold. In fact one can show that it consists of a finite number of points. Denote by kr(x−, x+) ∈ Z2 the parity of this number. Define a linear operator

∂r : Cm(F ) → Cm−1(F ) as follows. For every x ∈ Critm(F ) put X ∂rx = kr(x, y)y.

y∈Critm−1(F )

2 Fact 13.2.B. The operator ∂r is a differential: ∂r = 0. Thus (C(F ), ∂r) is a complex.

Fact 13.2.C. The m-th homology group Hm(C(F ), ∂r) of this complex is isomorphic to the homology Hm(M, Z2) of the manifold. In particular, though these homology groups depend on additional parameters F and r all of them are mutually isomorphic. A remarkable fact is that the isomorphisms can be arranged to a canonical family as follows. Consider the space of pairs (F, r) where F is a function and r is a Riemannian metric. Choose two generic pairs α = (F0, r0) and β = (F1, r1). Choose a generic path (Fs, rs), s ∈ R such that (Fs, rs) ≡ (F0, r0) for s ≤ 0 and (Fs, rs) ≡ (F1, r1) for s ≥ 1. Consider the equation 13.2. MORSE HOMOLOGY IN FINITE DIMENSIONS 135

du (s) = −∇ F (u(s)). (13.2.D) ds rs s

Let x− ∈ CritF0 and x+ ∈ CritF1 be two critical points. As before, for generic choices the space of solutions u(s) which satisfy u(s) → x− as s → −∞ and u(s) → x+ as s → +∞ is a smooth manifold of dimension i(x−) − i(x+). Further, when i(x−) = i(x+) there is only a 4 finite number of solutions. Denote by b(x−, x+) ∈ Z2 the parity of β,α this number. Define a linear operator I : C∗(F0) → C∗(F1) by X Iβ,α(x) = b(x, y)y. i(y)=i(x)

Fact 13.2.E.

• Every operator Iβ,α is a chain map and defines an isomorphism

β,α I∗ : H∗(C(F0), ∂r0 ) → H∗(C(F1), ∂r1 ).

β,α • The operator I∗ does not depend on the choice of the generic path (Fs, rs).

α,α γ,β β,α γ,α • I∗ = 1l and I∗ ◦ I∗ = I∗ for generic α, β, γ.

α,β We call a family of operators {I∗ } which satisfies the last property a canonical family.

Definition 13.2.F. Let (C, ∂) be a complex over Z2 with a given basis B = {e1, ..., ek}. An element e ∈ B is called homologically essential if for every ∂-invariant subspace

K ⊂ Span(B \{e}) the induced map H∗(K, ∂) → H∗(C, ∂)

4The crucial difference between equation 13.2.D and the gradient flow is that its solution space does not admit an R+-action when the family (Fs, rs) depends non-trivially on s. 136 CHAPTER 13. FLOER HOMOLOGY

is not surjective. The following statement plays a crucial role in our future considera- tions. Let F be a generic Morse function. Assume that x+ ∈ M is its unique point of absolute maximum. For a generic Riemannian metric r on M consider the complex (C(F ), ∂r) with the basis CritF .

Proposition 13.2.G. The point x+ ∈ CritF is homologically essen- tial. Sketch of the proof: Since the function F decreases along the tra- jectories of the negative gradient flow, the space

Q = Span(CritF \{x+}) is ∂r -invariant. Moreover HN (Q, ∂r) vanishes, where N = dimM. This reflects the fact that the manifold {F < a} is open for a ≤ max F and thus carries no fundamental class. We conclude that for every ∂-invariant subspace of Q the image of its N-th homology group in HN (C(F ), ∂r) vanishes. But

HN (C(F ), ∂r) = HN (M, Z2) = Z2.

Hence x+ is homologically essential. 2 This proposition yields the following important property of the coef- ficients b(x, y) which count (mod 2) solutions of 13.2.D. Let F be a generic Morse function with the unique absolute maximum x+. Con- sider a generic family (Fs, rs), s ∈ R as in 13.2.D such that Fs equals some Morse function F0 for all s ≤ 0 and Fs ≡ F for all s ≥ 1.

Corollary 13.2.H. There exists x ∈ CritF0 such that b(x, x+) 6= 0. Proof: Consider the operator

I :(C(F0), ∂r0 ) → (C(F ), ∂r1 ) defined by our data. If all b(x, x+) vanish, then the image of I is contained in Span(CritF \{x+}). This is a ∂r1 - invariant subspace.

Moreover, its homology coincides with H(C(F ), ∂r1 ) since I∗ is an isomorphism. We get a contradiction with the fact that x+ is homo- logically essential. 2 13.3. FLOER HOMOLOGY 137 13.3 Floer homology

The theory described in the previous section essentially generalizes to the infinite-dimensional situation. We replace the manifold M by its space of contractible loops LM, and the function space C∞(M) by the space of symplectic action functionals AF where F ∈ F. Let us start with the description of critical points of AF . Denote by P(F ) ⊂ LM the set of contractible 1-periodic orbits of the Hamil- tonian flow {ft} generated by F .

Exercise 13.3.A. Prove that the critical points of AF are exactly the elements of P(F ). Moreover, if the time-1 map f1 is non-degenerate in the sense that its graph is transversal to the diagonal in M × M, then every critical point of AF is non-degenerate. Denote by J the space of all Ω-compatible almost complex struc- tures on M. Choose an element J ∈ C∞(S1, J ). Every such J defines a Riemannian metric on LM in the following way. The tangent space to LM at a loop z ∈ LM consists of vector fields on M along z. Given two such vector fields, say ξ(t) and η(t) define their scalar product by

Z 1 Ω(ξ(t),J(t)η(t))dt. 0

Given J ∈ J we write ∇J for the gradient with respect to the Rie- mannian metric Ω(ξ, Jη) on M.

Exercise 13.3.B. Prove that the gradient of AF with respect to the Riemannian metric on LM described above is given by

du gradA (u) = J (t) + ∇ F (u(t), t). F dt J(t)

Thus in order to define Morse complex of AF one has to investigate the following problem. Find a smooth map u : R(s) × S1(t) → M such that

∂u ∂u (s, t) + J(t) (s, t) + ∇ F (u(s, t), t) = 0 (13.3.C) ∂s ∂t J(t) 138 CHAPTER 13. FLOER HOMOLOGY

and u(s, t) → z±(t) as s → ±∞, where z± ∈ P(F ). Denote by MF,J (z−, z+) the space of solutions of 13.3.C. Floer established that it has a very nice structure which we are going to describe now. The crucial point is that 13.3.C is a Fredholm problem. In fact we see that up to the 0-order terms this is just the familiar Cauchy-Riemann equation. It turns out that for generic F and J the space MF,J (z−, z+) is a smooth manifold with respect to a natural topology. Further, the dimension d(z−, z+) of this manifold alters nei- ther under homotopies of the Hamiltonian path {ft}, t ∈ [0; 1] with fixed end points, nor under changes of J. Thus this number depends only on the lift of the map {f1} to the universal cover of Ham(M, Ω). We arrived at one of the most surprising points of the whole the- ory. Notice that finite-dimensional intuition suggests that the number d(z−, z+) is simply the difference of the Morse coindices of AF at criti- cal points z− and z+. However it is very easy to see that these coindices are infinite. Nevertheless their difference makes sense! The general for- mula for d(z−, z+) is quite complicated — it requires the notion of the Conley-Zehnder index which we will not introduce in the book. How- ever the situation simplifies drastically for the case when F ∈ F(gε), where {gt} is a given Hamiltonian flow generated by a normalized time-independent Morse function G, and ε > 0 is small enough. From now on we describe Floer homology for Hamiltonians F ∈ F(gε). Exercise 13.3.D.

• Show that for ε > 0 small enough every ε-periodic orbit of the autonomous flow {gt} is a fixed point of the flow. Thus P(εG) coincides with the set CritG of the critical points of G.

• Prove that for F ∈ F(gε) every 1-periodic orbit of the flow {ft} is contractible. Moreover, the map

j : P(F ) → CritG, z(t) → z(0)

establishes between 1-periodic orbits of {ft} and critical points of G.

Denote by i(x) the Morse coindex of a critical point x ∈ CritG. 13.3. FLOER HOMOLOGY 139

Fact 13.3.E (The dimension formula). For a generic F ∈ F(gε), and for all z± ∈ P(F ) the following equality holds:

d(z−, z+) = i(z−(0)) − i(z+(0)).

Note that the solution space MF (z−, z+) admits a natural R- action by shifts u(s, t) → u(s + c, t) where c ∈ R. The boundary conditions guarantee that this action is free. Assume in addition that d(z−, z+) = 1. Then the dimension formula yields that the quotient space MF,J (z−, z+)/R is a zero dimensional manifold. A version of Gromov’s compactness theorem for solutions of 13.3.C states that this manifold is compact, hence consists of a finite number of points. De- note by kJ (z−, z+) ∈ Z2 the parity of this number. We proceed further exactly as in the finite-dimensional situation. Recall that critical points of AF are identified with CritG via the map j, see 13.3.D. Set Cm(AF ) = Cm(G) and C(AF ) = C(G). Fix a loop J(t) and define a linear operator ∂J : C(AF ) → C(AF ) as follows. For an element x ∈ CritmG set X −1 −1 ∂J (x) = kJ (j x, j y)y.

y∈Critm−1G

Fact 13.3.F. For generic F and J the operator ∂J is a differential : 2 ∂J = 0. Thus (C(AF ), ∂J ) is a chain complex. Example 13.3.G. Consider equation 13.3.C for F = εG and for a constant loop J(t) ≡ J. Every solution u(s) of the gradient flow equation du (s) = −ε∇ G(u(s)) ds J solves 13.3.C. A fairly involved argument [HS, Lemma 7.1] shows that for ε > 0 small enough these are the only solutions of the equation provided G and J are generic. Therefore (C(AF ), ∂J ) coincides with the usual Morse complex of the function F ! The following discussion emphasizes an important and delicate point in Floer theory. Let us try to prove the statement above for the case when d(z−, z+) = 1. We have to show that every solution 140 CHAPTER 13. FLOER HOMOLOGY

of equation 13.3.C is time independent. Since G is time indepen- 1 dent, the solution space MεG,J (z−, z+) admits a natural R×S -action 1 u(s, t) → u(s + c1, s + c2) where (c1, c2) ∈ R × S . Assume that u(s, t) is a solution which depends non-trivially on t. Then the R × S1-action is free in a neighbourhood of u, and thus the dimension d(z−, z+) is at least 2. This contradiction proves that u cannot depend on t. Un- fortunately, such an argument contains a serious gap. Namely, time- independent functions form a ” very thin” set in F, and thus no general position type argument can guarantee a priori that the space of so- lutions MεG,J (z−, z+) is a manifold! We refer to [HS] for a complete proof.

Let F0 ∈ F(gδ) and F1 ∈ F(gε) be two generic functions. Con- ∞ 1 sider a generic family (Fs,Js), s ∈ R where Fs ∈ F, Js ∈ C (S , J ) and (Fs,Js) ≡ (F0,J0) for s ≤ 0 and (Fs,Js) ≡ (F1,J1) for s ≥ 1. Ex- actly as in the finite dimensional situation we shall define a canonical homomorphism

I :(C(AF0 ), ∂J0 ) → (C(AF1 ), ∂J1 ). The following problem is analogous to 13.2.D. Find a smooth map u : R(s) × S1(t) → M such that ∂u ∂u (s, t) + J (t) (s, t) + ∇ F (u(s, t), t) = 0 (13.3.H) ∂s s ∂t Js(t) s and u(s, t) → z±(t) as s → ±∞, where z− ∈ P(F0) and z+ ∈ P(F1). The analysis of this equation is quite similar to the one we sketched above. In particular, if i(z−) = i(z+) then under certain genericity assumptions the space of solutions is zero dimensional and compact. Thus it consists of a finite number of points. Denote by b(z−, z+) ∈ Z2 the parity of this number. In contrast to the case of 13.3.C the solution space of 13.3.H is not invariant under translations u(s, t) → u(s + c, t) provided (Fs,Js) depends non-trivially on the s-variable. Define now a I as follows. For x ∈ CritmG set X I(x) = b(j−1x, j−1y)y. (13.3.I)

y∈CritmG 13.4. AN APPLICATION TO GEODESICS 141

In complete analogy with 13.2.E one can show that I induces an iso- morphism I∗ on homology which does not depend on the choice of the generic path (Fs,Js). Moreover, exactly as in 13.2 the family of maps I∗ associated with different choices of parameters (F,J) can be ar- ranged into a canonical family. In particular, 13.3.G and 13.2.C above imply that for generic F and J

H∗(C(AF ), ∂J ) = H∗(M, Z2).

13.4 An application to geodesics

Now we have the necessary tools which will enable us to prove Theo- rem 12.6.F. Let G be a normalized Morse function on M with unique absolute maximum point x+ and unique absolute minimum point x−. Write {gt} for its Hamiltonian flow. Take any Hamiltonian path {ft}, t ∈ [0; 1] with f0 = 1l,f1 = gε such that {ft} is homotopic to {gεt}, t ∈ [0; 1] with fixed end points. Let F ∈ F be its normalized Hamiltonian. Proposition 13.4.A. The following inequalities hold:

Z 1 max F (x, t) ≥ ε max G, 0 x

Z 1 min F (x, t) ≤ ε min G. 0 x

Theorem 12.6.F immediately follows from this proposition. For the proof of 13.4.A we start with the following auxiliary state- −1 ∗ ment. Consider the loop ht = ft ◦ gεt . Note that AF = Th AεG in view of 13.1.A. Fix a generic time-independent almost complex structure −1 J0 ∈ J on M, and consider the loop J(t) = ht∗ J0ht∗. Denote by r0 and r Riemannian metrics on LM associated to J0 and J(t) respec- ∗ tively. Then r = Th r0. Thus with the help of Th one can identify the

complexes (C(AεG), ∂J0 ) and (C(AF ), ∂J ). Moreover this identification preserves the basis CritG of both complexes pointwise. 142 CHAPTER 13. FLOER HOMOLOGY

Lemma 13.4.B. The point x+ is homologically essential in (C(AF ), ∂J ).

Proof: Indeed, 13.3.G and 13.2.G yield that x+ is homologically es- sential in (C(AεG), ∂J0 ). The result immediately follows from the iden- tification above. 2

Proof of 13.4.A: Fix δ ∈ (0; ²). Let a(s), s ∈ R be a non-decreasing function such that a(s) ≡ 0 for s ≤ 0 and a(s) ≡ 1 for s ≥ 1. Set

Fs(x, t) = (1 − a(s))δG(x) + a(s)F (x).

1 Choose a generic path Js : S → J such that Js(t) ≡ J0 for s ≤ 0 and Js(t) ≡ J(t) for s ≥ 1. Consider the homomorphism I defined in 13.3.I. Set z+(t) = ftx+ ∈ P(F ). Using that x+ is homologically essential (see 13.4.B) and arguing exactly as 13.2.H we get that b(z−, z+) 6= 0 for some z− ∈ P(δG). Hence there exists a solution u(s, t) of the problem 13.3.H. Consider the energy integral Z Z µ ¶ +∞ 1 ∂u ∂u E = ds dt Ω (s, t),Js(t) (s, t) . −∞ 0 ∂s ∂s

Using equation 13.3.H and the fact that sgradF = J∇J F for every J ∈ J we compute µ ¶ µ ¶ µ ¶ ∂u ∂u ∂u ∂u ∂u Ω ,J (t) = Ω , − dF . (13.4.C) ∂s s ∂s ∂s ∂t s ∂s

Let D− and D+ be discs in M bounded by z− and z+ respectively. Since π2(M) = 0, we get that Z Z µ ¶ Z Z +∞ 1 ∂u ∂u ds dt Ω , = Ω − Ω. (13.4.D) −∞ 0 ∂s ∂t D+ D− Further, µ ¶ ∂u d ∂F dF = (F (u(s, t), t)) − s (u(s, t), t). (13.4.E) s ∂s ds s ∂s Integrating 13.4.C and substituting 13.4.D,E we get that

E = AδG(z−) − AF (z+) + Q, 13.5. TOWARDS THE EXIT 143

where Z Z +∞ 1 da Q = ds dt (s)(F (u(s, t), t) − δG(u(s, t))). −∞ 0 ds

Clearly Z 1 Q ≤ max(F (x, t) − δG(x))dt, 0 x

AF (z+) = AεG(x+) = ε max G, and E ≥ 0. Moreover, AδG(z−) ≤ δ max G since all closed orbits of the flow {gδt} are simply critical points of G. Therefore

Z 1 max(F (x, t) − δG(x)) ≥ (ε − δ) max G. 0 x Since this is true for all δ > 0 we get that

Z 1 max(F (x, t) ≥ ε max G. 0 x This proves the first inequality in 13.4.A. The second one can be proved analogously. 2

13.5 Towards the exit

We start with a short summary of the material presented in this chap- ter. Let (M, Ω) be an aspherical symplectic manifold. 13.5.A. To every generic function F ∈ F one associates a vector space C(AF ) over Z2. This space is endowed with a preferred basis P(F ) consisting of all contractible periodic orbits of the correspond- ing Hamiltonian flow. Also, C(AF ) has a Z-grading in terms of the Conley-Zehnder index. We described this grading explicitly in terms of the Morse index in the simplest case when F belongs to F(gε). 144 CHAPTER 13. FLOER HOMOLOGY

13.5.B. A generic choice of a loop of compatible almost complex struc- ∞ 1 tures J ∈ C (S , J ) defines a differential ∂J : C∗(AF ) → C∗−1(AF ). The homology of the complex (C(AF ), ∂J ) is isomorphic to H∗(M, Z2).

13.5.C. Given generic (F0,J0) and (F1,J1) and a generic path (Fs,Js) interpolating between them, there exists a canonical chain map I :

(C(AF0 ), ∂J0 ) → (C(AF1 ), ∂J1 ). The map I induces an isomorphism I∗ on homology which does not depend on the choice of the path. The isomorphisms I∗ form a canonical family in the sense of 13.2.E.

13.5.D. Assume that F0 and F1 generate the same element of Ham(] M, Ω). Then F0 = h(F1) for a loop h ∈ LHam(M, Ω) (see 13.1.A). ∞ 1 −1 Given J0(t) ∈ C (S , J ), set J1(t) = ht∗ J0(t)ht∗. Then the map

Th : LM → LM introduced in 13.1 identifies (C(AF0 ), P(F0), ∂J0 ) with (C(AF1 ), P(F1), ∂J1 ). This identification is much stronger than the one described in 13.5.C - it works on the chain level while the I’s induce an isomorphism on the homological level only. The structure described in 13.5.A – 13.5.D above is the simplest part of what is called Floer homology theory associated to a symplectic manifold. The next statement relates it to Hofer’s norm. 13.5.E. Assume that for some J ∈ C1(S1, J ) a contractible periodic orbit z ∈ P(F ) is homologically essential in (C(AF ), P(F ), ∂J ). Then

Z 1 AF (z) ≤ max F (x, t)dt. 0 x

The proof is absolutely analogous to the proof of 13.4.A presented in 13.4 above. Further, denote by φ ∈ Ham(] M, Ω) the element generated by F . In view of 13.5.D the property of an orbit z ∈ P(F ) to be homo- logically essential for some J is an intrinsic property of the fixed point z(0) of φ. The value AF (z) does not depend on the specific choice of F ∈ F(φ). Thus inequality 13.5.E provides a tool for estimating the Hofer distance between 1l and φ on the universal cover Ham(] M, Ω). It remains to develop a machinery which enables one to decide which orbits are homologically essential. In 13.4.B above we worked out the simplest situation when z corresponds to the absolute maximum of a small time-independent Hamiltonian function. In [Sch3] Schwarz used 13.5. TOWARDS THE EXIT 145 a more sophisticated argument which enabled him to prove the local minimality of a wide class of geodesics for aspherical symplectic mani- folds. In fact, there is an important additional piece of data associated canonically to (C(AF ), P(F )), namely the canonical real filtration of a C(AF ). For a ∈ R denote by C the subspace of C(AF ) generated by those z ∈ P(F ) which satisfy AF (z) ≤ a. Since the action func- tional decreases along the trajectories of its negative gradient flow, this subspace is ∂J -invariant. Thus one can speak about the relative a b homology groups H(C /C , ∂J ). It turns out that this homology car- ries a lot of interesting information about φ. Such a filtration was first systematically used by Viterbo [V1]. Further results on this filtration were obtained by Oh [O4] and Schwarz [Sch3] . Let me mention that there exist two natural directions of gen- eralizing the theory sketched above. The first one is to extend it to symplectic manifolds with non-trivial π2. For such manifolds the ac- tion functionals AF are multi-valued functions LM → R. In fact, their differentials dAF are well-defined closed 1-forms on LM. Floer homol- ogy in this situation is designed as a generalization of Morse-Novikov homology of closed 1-forms (see [HS]). The second direction is to extend cohomological operations (like the cup-product) to Floer ho- mology (see [PSS]). Let me mention finally that Floer homology is a very complicated structure associated to Ham(] M, Ω). An interesting task which is still far from being fulfilled is to develop an algebraic language suitable for a transparent description of this structure. We refer to [Fu] for first steps in this direction. 146 CHAPTER 13. FLOER HOMOLOGY Chapter 14

Geometry of non-Hamiltonian diffeomorphisms

In this chapter we discuss interrelations between the groups Ham(M, Ω)

and Symp0(M, Ω), and explain the role played by non-Hamiltonian diffeomorphisms in Hofer’s geometry.

14.1 The flux homomorphism

Let (M, Ω) be a closed symplectic manifold and let Symp0(M, Ω) be the identity component of the group of symplectomorphisms (see 1.4.C). Given a path {ft} of symplectomorphisms with f0 = 1l, con-

sider a vector field ξt on M which generates this path. Since Lξt Ω = 0

then iξt Ω = λt is a family of closed 1-forms. Let us emphasize that these forms are not necessarily exact. Let us consider a basic example which we started to discuss in the beginning of the book (see 1.4.C). Example 14.1.A. Consider the 2-torus (T2 = R2/Z2, dp ∧ dq) and the following system ½ p˙ = 0 q˙ = 1 corresponding to the path of symplectomorphisms ft(p, q) = (p, q + t).

147 148 CHAPTER 14. NON-HAMILTONIAN DIFFEOMORPHISMS

This path is generated by an autonomous family of closed 1-forms λt = dp. The first question we want to investigate is the following one. Take any f ∈ Symp0(M, Ω) (say given by a precise formula). How do we decide whether f is Hamiltonian or not? A powerful tool which will enable us to answer this question in the example above is the notion of the flux homomorphism introduced by Calabi and studied further by Banyaga [B1]. We will discuss it in this chapter. Note that in general one cannot conclude that f is not Hamiltonian when it is generated by a non-exact 1-form. For instance, in the above example f1 = 1l is Hamiltonian. Let us introduce the following useful notion. Let {ft}, t ∈ [0, 1] be a loop of symplectomorphisms with f0 = f1 = 1l. Let {λt} be the family of closed 1-forms generating this loop. R 1 Definition. The flux of a loop is given by flux({ft}) = 0 [λt]dt ∈ H1(M, R). Let us give a more geometric description of the flux. Let C be a 1-cycle on M. Define a 2-cycle ∂[C] = ∪ft(C) which is the image of t C under the flow of ft. Note that ∂ is a linear map from H1(M, Z) to H2(M, Z).

Exercise. Show that (flux({ft}), [C]) = ([Ω], ∂[C]) for all [C] ∈ H1(M, Z).

In particular, this shows that flux({ft}) depends only on the homotopy class of {ft} in π1(Symp0(M, Ω)), since the right hand side persists under homotopies. Thus we get a homomorphism

1 flux : π1(Symp0(M, Ω)) → H (M, R).

Definition. The image of the flux homomorphism Γ ⊂ H1(M, R) is called the flux group. If γ is represented by a Hamiltonian loop then flux(γ) = 0. A fundamental result due to Banyaga ([B1]) states that the converse is true. That is flux(γ) = 0 implies that γ can be homotoped to a Hamiltonian loop. 14.1. THE FLUX HOMOMORPHISM 149

It will be useful to extend the notion of flux to arbitrary (not nec- essarily closed) smooth paths of symplectomorphisms. Given a path {ft} generated by a family of closed 1-forms λt we put Z 1 1 flux({ft}) = [λt]dt ∈ H (M, R). 0

Take a symplectomorphism f ∈ Symp0(M, Ω) and choose any path {ft} with f0 = 1l and f1 = f. Clearly, flux({ft}) depends on the choice of the path connecting 1l with f, but the difference between the fluxes of any two such paths belongs to Γ. Thus we get a mapping 1 ∆ : Symp0(M, Ω) → H (M, R)/Γ. We leave it as an exercise to check that ∆ is a homomorphism. Banyaga showed that ker(∆) = Ham(M, Ω), thus 1 Symp0(M, Ω)/Ham(M, Ω) = H (M, R)/Γ. Exercise 14.1.B. Combine this with Theorem 1.5.A above and show −1 that every normal subgroup of Symp0(M, Ω) has the form ∆ (K) for some subgroup K of H1(M, R)/Γ. Let us calculate the flux group for (T2, dp ∧ dq) (see 14.1.A above). Make the following identifications: 2 2 2 1 2 2 2 1 2 H1(T , Z) = Z ,H2(T , Z) = Z,H (T , Z) = Z ⊂ R = H (T , R). We claim that with this language Γ = Z2. Indeed, for each γ ∈ 2 π1(Symp0(T )) we have (flux(γ), a) = ([dp ∧ dq], ∂a). Here ∂ is the functional which takes Z2 to Z, the value of the class 2 2 [dp ∧ dq] on the generator of H2(T , Z) equals 1, and a ∈ H1(T , Z). Thus flux(γ) ∈ H1(T2, Z) which implies that Γ ⊂ Z2. On the other hand, the fluxes of full rotations of the torus are given by ¡ ¢ flux (p, q) 7→ (p, q + t) = [dp] , ¡ ¢ flux (p, q) 7→ (p + t, q) = −[dq] , and we conclude that Z2 ⊂ Γ. Therefore Γ = Z2. In particular, we learn that fT (p, q) = (p, q + T ) is Hamiltonian precisely when T ∈ Z. 2 Indeed, ∆(fT ) = T [dp] mod Z . 150 CHAPTER 14. NON-HAMILTONIAN DIFFEOMORPHISMS 14.2 The flux conjecture

In general, it is not so easy to calculate the flux group. The simple question whether Γ is discrete turns out to be an important one by the following reason. Recall (see 1.4.F above) that the flux conjecture1 ∞ states that Ham(M, Ω) is C -closed in Symp0(M, Ω). Theorem 14.2.A. If Γ is discrete then the flux conjecture holds.

ε path with small flux k

φ k .. φ

Hamiltonian path Flux = 0

Flux = λ . 1 Figure 13

∞ Proof: Let φk be a sequence of Hamiltonian diffeomorphisms C - converging to φ ∈ Symp0(M, Ω). Fix a path of symplectomorphisms which joins φ with the identity. Write λ ∈ H1(M, R) for its flux. We claim that there exists a sequence of paths connecting φk to φ whose fluxes εk converge to 0 when k goes to +∞. Assume the claim. Then (see figure 13) for each k there exists a loop with flux λ+εk ∈ Γ. Since Γ is discrete we conclude that εk must equal zero for large enough k. Thus φ is Hamiltonian since ∆(φ) = 0 mod Γ, so the theorem follows from the claim. It remains to prove the claim. It suffices to find a sequence of −1 paths γk connecting 1l with φk φ whose fluxes converge to 0.

1This is the C∞-flux conjecture. Note that it is equivalent to the C1-flux con- jecture. In contrast to this, the C0-flux conjecture is a much stronger statement (see [LMP1] for discussion). 14.2. THE FLUX CONJECTURE 151

graph( φ -1 φ) k ∆ (M,- Ω)

(M,Ω) Figure 14

−1 −1 Note that φk φ → 1l when k → +∞, so graph (φk φ) is close to the diagonal L in (M × M, Ω ⊕ −Ω). Now we use the following trick. Recall that L is a Lagrangian submanifold in (M × M, Ω ⊕ −Ω).

Lemma. Let (P, ω) be a symplectic manifold, and let L ⊂ P be a closed Lagrangian submanifold. There exists a neighbourhood U of L in P and an embedding f : U → T ∗L with the following properties:

∗ ∗ • f ω0 = ω, where ω0 is the standard symplectic structure on T L; • f(x) = (x, 0) for all x ∈ L.

This is a version of the Darboux theorem for Lagrangian submanifolds [MS]. Let us apply this lemma to the diagonal L ⊂ M × M. It allows us to identify a tubular neighbourhood of L with a tubular neigbourhood of the zero section in T ∗L. Since T ∗L ' T ∗M, we see that for large −1 ∗ k the graph of (φk φ) corresponds to a section of T M that is to a closed 1-form αk on M. Clearly αk → 0 when k → +∞. Fix k large enough and consider the deformation of Lagrangian submanifolds

∗ graph (sαk) ⊂ T M

for s ∈ [0, 1]. For every s the submanifold graph (sαk) is identified with the graph of a symplectomorphism of M. Denote this symplecto- morphism by γk(s). Clearly the flux of the path γk(s) equals [αk]. This 152 CHAPTER 14. NON-HAMILTONIAN DIFFEOMORPHISMS

provides us with the required path, and hence completes the proof of the theorem. 2

In the proof above the main difficulty of the flux conjecture be- comes especially transparent. It lies in the fact that for an arbitrary Hamiltonian isotopy {ft} , the graph of {ft} may leave the tubular neighbourhood of L for some t. It will come back to this neighbour- hood, but the question is whether it is then still the graph of an exact 1-form. Note also that the converse of the above theorem is also true (see [LMP1]).

Corollary. If [Ω] ∈ H2(M, Q) then the flux conjecture holds.

1 2 Proof: We can rewrite the assumption as [Ω] ∈ k H (M, Z) for some k ∈ Z. So 1 (flux(γ), [C]) = ([Ω], ∂[C]) ∈ Z k 1 1 1 for every γ ∈ π1(Symp0(M, Ω)). Therefore Γ ⊂ k H (M, Z) ⊂ H (M, R) and we deduce that Γ is discrete. Now the above theorem implies that the flux conjecture holds. 2

14.3 Links to “hard” symplectic topol- ogy

Here is a more conceptual proof of the flux conjecture for T2. Clearly k the C -closure of Ham(M, Ω) in Symp0(M, Ω) is a normal subgroup of Symp0(M, Ω). Hence in view of 14.1.B above it suffices to show 1 2 that for every α ∈ H (T , R)/Γ \{0} there exists f ∈ Symp0(M, Ω) with ∆(f) = α which cannot be represented as a limit of Hamilto- nian diffeomorphisms. Without loss of generality take f to be a shift (p, q) 7→ (p, q + T ), where T/∈ Z. The famous Arnold conjecture proved in [CZ] states that every φ ∈ Ham(T 2) has a fixed point. Thus if φk → f, this would imply that f has a fixed point too, which is a contradiction. This argument is due to M. Herman (1983), and it works for T2n as well. Note also, that it proves the C0-flux conjecture. 14.4. ISOMETRIES IN HOFER’S GEOMETRY 153

One can try to generalize this argument to other symplectic mani- folds. The idea is to consider instead of the limiting behaviour of fixed points the limiting behaviour of Floer homology. In this way, one gets the following result.

Theorem 14.3.A. ([LMP1]) Assume that the first Chern class c1(TM) vanishes on π2(M). Then the flux conjecture holds.

Here is another application of the theory of pseudo-holomorphic curves to the flux conjecture, which was found in [LMP2] for 4-dimensional symplectic manifolds, and later on proved in [McD2] in full generality.

Theorem 14.3.B. The rank of the flux group Γ is finite and satisfies

1 rankZ Γ ≤ b1(M) = dim H (M, R).

As a corollary we get that for symplectic manifolds whose first Betti number equals 1, the group Γ is discrete, and therefore the flux con- jecture holds.

14.4 Isometries in Hofer’s geometry

No natural metric on Symp0(M, Ω) is known. However, one can in- clude general non-Hamiltonian symplectomorphisms into the frame- work of Hofer’s geometry as follows (see [LP]) . Observe that the group Symp(M, Ω) of all symplectomorphisms of (M, Ω) acts on Ham(M, Ω) by isometries. For φ ∈ Symp(M, Ω), define

−1 Tφ : Ham(M, Ω) → Ham(M, Ω) , f 7→ φfφ .

One can easily check that Tφ is well defined and that it is an with respect to Hofer’s metric ρ. Definition. An isometry T is bounded if sup ρ(f, T f) < ∞, where the infimum is taken over all f ∈ Ham(M, Ω). 154 CHAPTER 14. NON-HAMILTONIAN DIFFEOMORPHISMS

If φ is Hamiltonian, then the isometry Tφ is bounded. Indeed, the following inequality gives us a universal bound independent of f:

−1 −1 −1 ρ(f, Tφf) = ρ(f, φfφ ) = ρ(1l, φfφ f ) ≤ 2ρ(1l, φ).

Define the set BI0 ⊂ Symp0(M, Ω) as the set of all φ ∈ Symp0(M, Ω) such that Tφ is bounded. As we have seen above Ham(M, Ω) ⊂ BI0.

Exercise 14.4.A. Prove that BI0 is a normal subgroup of Symp0(M, Ω). The next conjecture suggests a characterization of Hamiltonian diffeo- morphisms in metric terms.

Conjecture. Ham(M, Ω) = BI0.

Theorem 14.4.B. ([LP]) The conjecture is true for surfaces of genus ≥ 1 and their products. We will give the idea of the proof for the case when M is the 2-torus T2. It follows from 14.4.A that in order to prove the theorem it suffices to show the following. Given

(a, b) ∈ H1(T2; R) \ Γ = R2 \ Z2,

2 there exists a symplectomorphism φ ∈ Symp0(T ) with ∆(φ) = (a, b) 2 mod Z such that the corresponding isometry Tφ is unbounded. As- sume without loss of generality that a = 0, b ∈ (0; 1) and φ(p, q) = (p, q + b). Note that for a curve C = {q = 0} we have C ∩ φ(C) = ∅. Let F = F (q) be a normalized Hamiltonian on T2 whose support lies in a small neigbourhood of C and such that F |C ≡ 1. Denote by ft the corresponding Hamiltonian flow, and consider the flow formed by the −1 −1 commutators gt = φ ft φft. This flow is generated by the Hamil- tonian G(q) = F (q) − F (q + b). Since G|C ≡ 1 we get from 7.4.A that ρ(1l, gt) goes to infinity when t → ∞. Thus the isometry Tφ is unbounded. The problem of describing all isometries of (Ham(M, Ω), ρ) is open and seems to be very difficult even for surfaces. In connection to this let me recall the following classical result due to Mazur-Ulam [MU] (1932): every 0-preserving surjective isometry of a linear normed 14.4. ISOMETRIES IN HOFER’S GEOMETRY 155

space is a linear map. It would be interesting to prove, or disprove, a non-linear version of this statement: every 1l-preserving isometry of Ham(M, Ω) is a group isomorphism (maybe after composition with the involution f 7→ f −1). Assume for a moment that this was indeed true. Then Banyaga’s theorem 1.5.D would yield that each such isometry (up to the abovementioned involution) equals Tφ, where φ : M → M is either a symplectomorphism, or an anti-symplectomorphism (that is φ∗Ω = −Ω). This would mean that Hofer’s geometry determines the symplectic topology. 156 CHAPTER 14. NON-HAMILTONIAN DIFFEOMORPHISMS Bibliography

[A] M.Abreu Topology of symplectomorphism groups of S2 × S2, In- vent. Math. 131 (1998), 1-23.

[AM] M.Abreu & D.McDuff Topology of symplectomorphism groups of rational ruled surfaces, Preprint math/9910057, 1999.

[AS] L.Ahlfors & L.Sario Riemann Surfaces, Princeton University Press, 1960.

[Ar] V.Arnold Mathematical methods in Classical Mechanics, Springer, Berlin, 1978.

[AK] V.Arnold & B.Khesin Topological methods in hydrodynamics Springer, 1998.

[AL] M.Audin & J.Lafontaine Holomorphic curves in symplectic ge- ometry, Progress in mathematics 117, Birkh¨auser,1994.

[B1] A.Banyaga Sur la structure du groupe des diff´eomorphismesqui pr´eserventune forme symplectique, Comm.Math.Helv. 53 (1978), 174-227.

[B2] A.Banyaga The structure of classical diffeomorphism groups, Mathematics and its applications 400, Kluwer Academic Pub- lisher’s Group, 1997.

[BP1] M.Bialy & L.Polterovich Geodesics of Hofer’s metric on the group of Hamiltonian diffeomorphisms, Duke Math. J. 76 (1994), 273-292.

157 158 BIBLIOGRAPHY

[BP2] M.Bialy & L.Polterovich Invariant tori and symplectic topology, Amer. Math. Soc. Transl. 171 (1996), 23-33. [Ch] Y.Chekanov Lagrangian intersections, symplectic energy, and areas of holomorphic curves, Duke Math. J. 95 (1998), 213 -226. [CFS] I.Cornfeld, S.Fomin & Y.Sinai Ergodic Theory, Springer, 1982. [CZ] C.Conley & E.Zehnder The Birkhoff-Lewis fixed point theorem and a conjecture of V.I.Arnold, Invent. Math. 73 (1983), 33-49. [DFN] B.A.Dubrovin, A.T.Fomenko & S.P.Novikov Modern Geometry — methods and applications, Part I, Springer, 1984. [EE] C. Earle and J. Eells, A fibre bundle description of Teichm¨uller theory, J. Differential Geometry 3(1969)19-43. [EP] Y.Eliashberg & L.Polterovich Bi-invariant metrics on the group of Hamiltonian diffeomorphisms, Internat.J.Math. 4 (1993), 727- 738. [F] A.Floer Morse theory for Lagrangian intersection theory, J.Diff.Geometry 28 (1988), 513-517. [Fu] K.Fukaya Morse homotopy and its quantization Geometric topol- ogy (Athens, GA, 1993), 409–440, AMS/IP Stud. Adv. Math., 2.1, Amer. Math. Soc., Providence, RI, 1997. [GH] P.Griffiths & J.Harris Principles of algebraic geometry, Wiley, New York, 1978. [G1] M.Gromov Pseudoholomorphic curves in symplectic manifolds, Invent.Math. 82 (1985), 307-347. [G2] M.Gromov Positive curvature, macroscopic dimension, spectral gaps and higher signatures, in ”Functional Analysis on the eve of the 21st century”, S. Gindikin, J. Lepowsky, R. Wilson eds., Birkh¨auser,1996. [GLS] V.Guillemin, E.Lerman & S.Sternberg Symplectic fibrations and multiplicity diagrams, Cambridge University Press, 1996. BIBLIOGRAPHY 159

[He] M.Herman Inegalites a priori pour des tores lagrangiens invari- ants par des diffeomorphismes symplectiques, Publ. Math. IHES 70(1989), 47-101.

[H1] H.Hofer On the topological properties of symplectic maps, Proc. Royal Soc. Edinburgh 115A (1990), 25-38.

[H2] H.Hofer Estimates for the energy of a symplectic map, Commen- tarii Math. Helv. 68(1993), 48-72.

[HS] H.Hofer & D.Salamon Floer Homology and Novikov rings, in “Volume in Memory of Andreas Floer” (H. Hofer, C. Taubes, A. Weinstein, E. Zehnder, eds.), Progress in Mathematics, Birkh¨auser,1995, pp. 483-524.

[HZ] H.Hofer & E.Zehnder Symplectic invariants and Hamiltonian dy- namics, Birkh¨auserAdvanced Texts, Birkh¨auserVerlag, 1994.

[Ki] Y.Kifer Random dynamics and its applications, in ”Proceedings of the ICM, Berlin, 1998”, Vol. II, pp. 809-818.

[L] F.Lalonde Energy and capacities in symplectic topology, AMS/IP Studies in Advanced Mathematics, Vol.2.1 (1997), 328-374

[LM1] F.Lalonde & D.McDuff The geometry of symplectic energy, Ann. of Math. 141 (1995), 349-371.

[LM2] F.Lalonde & D.McDuff Hofer’s L∞-geometry: energy and sta- bility of Hamiltonian flows I, II, Invent. Math 122 (1995), 1-33 , 35-69.

[LM3] F.Lalonde & D.McDuff Local non-squeezing theorems and sta- bility, Geom. and Funct. Anal. 5(1995), 364-386.

[LMP1] F.Lalonde, D.McDuff & L.Polterovich On the flux , ”Geometry, Topology and Dynamics, ed F. Lalonde, Proceedings of the CRM 1995 Workshop in Montreal, the CRM Special Series of the AMS”. Vol.15, 1998, pp. 69-85. 160 BIBLIOGRAPHY

[LMP2] F.Lalonde, D.McDuff & L.Polterovich Topological rigidity of Hamiltonian loops and quantum homology, Invent. Math. 135 (1999), 69-85.

[LP] F.Lalonde & L.Polterovich Symplectic diffeomorphisms as isome- tries of Hofer’s norm, Topology 36 (1997), 711-727.

[LS] F.Laudenbach & J.-C.Sikorav Hamiltonian disjunction and limits of Lagrangian submanifolds, Internat. Math. Res. Notices 1994, no.4 161-168.

[Mac] R.MacKay A criterion for nonexistence of invariant tori for Hamiltonian systems, Phys. D 36 (1989), 64-82.

[MU] S.Mazur & S.M.Ulam Sur les transformations isometriques d’espaces vectoriels normes C.R. Acad. Sci. Paris 194(1932), 946- 948.

[MS] D.McDuff & D.Salamon Introduction to symplectic topology, Ox- ford Mathematical Monographs, Oxford University Press, 1995.

[McD1] D.McDuff, Rational and ruled symplectic 4-manifolds Journal of the A.M.S. 3 (1990) 679-712; erratum 5 (1992), 987-988.

[McD2] D.McDuff Quantum homology of fibrations over S2, Preprint math/9905092, 1999, to appear in International Journal of Math- ematics.

[NN] A.Newlander & L.Nirenberg Complex analytic coordinates in al- most complex manifolds, Ann. Math. 65 (1957), 391-404.

[O1] Y.-G.Oh Floer cohomology of Lagrangian intersections and pseudo-holomorphic disks I, II, Comm.Pure Appl.Math. 46 (1993), 949-993, 995-1012.

[O2] Y.-G.Oh Relative Floer and quantum cohomology and the sym- plectic topology of Lagrangian submanifolds, in Contact and sym- plectic geometry, C.B.Thomas ed., Cambridge University Press, 1996, pp. 201-267. BIBLIOGRAPHY 161

[O3] Y.-G.Oh Gromov-Floer theory and disjunction energy of compact Lagrangian embeddings, Math. Res. Lett. 4 (1997), 895-905.

[O4] Y.-G. Oh Symplectic topology as the geometry of action func- tional - I,II, J. Differential Geometry 46 (1997), 499-577 and Comm. in Analysis and Geometry 7 (1999), 1-55.

[P1] L.Polterovich Symplectic displacement energy for Lagrangian submanifolds, Ergod. Th. & Dynam. Sys. 13 (1993), 357-367.

[P2] L.Polterovich An obstacle to non-Lagrangian intersections, In “Volume in Memory of Andreas Floer” (H. Hofer, C. Taubes, A. Weinstein, E. Zehnder, eds.), Progress in Mathematics, Birkh¨auser,1995, pp. 575-586.

[P3] L.Polterovich Gromov’s K-area and symplectic rigidity, Geom.Funct.Anal. 6 (1996), 726-739.

[P4] L.Polterovich Symplectic aspects of the first eigenvalue , J. reine angew. Math. (Crelle’s journal) 502 (1998), 1-17.

[P5] L.Polterovich Hofer’s diameter and Lagrangian intersections, Int.Math.Research Notes 4 (1998), 217-223.

[P6] L.Polterovich Hamiltonian loops and Arnold’s principle, Amer. Math. Soc. Transl. (2), 180 (1997), 181-187.

[P7] L. Polterovich Precise measurements in symplectic topology, in ”Proceedings of the 2nd European congress of Mathematics”, Vol. 2, Birkh¨auser,1998, pp. 159-166.

[P8] L. Polterovich Geometry on the group of Hamiltonian diffeomor- phisms, in ”Proceedings of the ICM, Berlin, 1998”, Vol. II, pp. 401-410.

[P9] L. Polterovich Hamiltonian loops from the ergodic point of view, Journal of the European Math. Soc. 1 (1999), 87 - 107.

[PS] L. Polterovich & K.F. Siburg On the asymptotic geometry of area- preserving maps, to appear in Math. Research Letters. 162 BIBLIOGRAPHY

[PSS] S.Piunikhin, D.Salamon & M.Schwarz Symplectic Floer- Donaldson theory and quantum cohomology, In ”Contact and symplectic geometry (Cambridge, 1994)”, C.B.Thomas ed., Publ. Newton Inst., 8, Cambridge Univ. Press, Cambridge, 1996, pp. 171-200.

[Sch1] M.Schwarz Morse homology, Birkh¨auser,1993.

[Sch2] M.Schwarz Introduction to symplectic Floer homology, In ”Con- tact and symplectic geometry (Cambridge, 1994)”, C.B.Thomas ed., Publ. Newton Inst., 8, Cambridge Univ. Press, Cambridge, 1996, pp. 151-170.

[Sch3] M.Schwarz On the action spectrum for closed symplectically as- pherical manifolds, Pacific J. of Math. 193 (2000), 419-461.

[Se] P.Seidel π1 of symplectic groups and invertibles in quantum homology rings, Geom. and Funct. Anal. 7(1997), 1046 -1095.

[S1] J.-C.Sikorav Quelques propri´et´es des plongements lagrangiens, Mm. Soc. Math. France (N.S.) 46 (1991), 151–167.

[S2] J.-C.Sikorav Syst`emes Hamiltoniens et topologie symplectique, ETS EDITRICE PISA, 1990.

[Si1] K.F.Siburg New minimal geodesics in the group of symplectic diffeomorphisms, Calc.Var. 3 (1995),299-309.

[Si2] K.F.Siburg Action-minimizing measures and the geometry of the Hamiltonian diffeomorphism group, Duke Math. J. 92 (1998),295- 319

[Sli] J.Slimowitz Hofer-Zehnder capacity and length minimizing paths in the Hofer norm, Preprint math/990515, 1999.

[Sm] S.Smale An infinite dimensional version of Sard’s theorem, Am. J. Math. 87(1973), 213-221. BIBLIOGRAPHY 163

[U] I.Ustilovsky Conjugate points on geodesics of Hofer’s metric, Diff.Geometry Appl. 6 (1996), 327-342.

[V1] C.Viterbo Symplectic geometry as the geometry of generating functions, Math. Annalen 292 (1992), 685-710.

[V2] C.Viterbo Metric and isoperimetric problems in symplectic ge- ometry, Journal of AMS 13 (2000), 411-431. Index

Abreu, 62, 105 Darboux theorem, 5, 151 adjoint action, 11 displacement energy, 19 Akveld, v of Lagrangian submanifolds, almost complex structure, 94 28 Arnold, 2, 70, 152 asymptotic growth, 72 Eliashberg, 18, 20 exact Lagrangian isotopy, 53 Banyaga, 8, 13, 14, 148, 155 exceptional sphere, 94 Bialy, 71, 72, 126, 128 biinvariant (pseudo)distance, 17 fixed extrema, 114 bounded isometry, 153 flatness phenomenon, 128 Floer, 57, 58, 63, 133 Calabi, 148 Floer homology, 131 canonical coordinates, 5 flow, 5 Chekanov, 18, 30 flux closed orbits, 127 conjecture, 10 compactness theorem, 38, 99 group, 148 compatible, 95 of a loop, 148 Conley, 152 Fukaya, 145 coupling class, 85 geodesic, 113 construction, 85 minimal, 80 form, 85 locally, 126 curve shortening, 76, 106 non-degenerate, 116 cusp stable, 116 curve, 100 strictly minimal, 72 solution, 38 locally, 125 cut off, 20 Gromov, 18, 28, 38, 57, 62, 76, cylindrical capacity, 29 86, 95, 99

164 INDEX 165

Guillemin, 85 theorem, 2 Hamiltonian Marsden, 82 diffeomorphism, 7 Mather, 76 equation, 6 Mazur-Ulam theorem, 155 flow, 7 McDuff, 18, 20, 62, 65, 80, 105, function, 6 108, 115–117, 126–128, loop, 46 153 path, 8 mechanical motion, 2 system, 1 Morse homology, 133 vector field, 5 Moser, 63, 70, 74, 82, 97 Herman, 70, 152 Hofer, i, 18–20, 29, 72, 128, 139, non-squeezing, 29 145 normalized Hamiltonian, 6 Hofer’s metric, 23 Novikov, 145 non degeneracy of, 29 Oh, 18, 30, 58, 145 Hofer’s norm, 23 homologically essential, 135 path of diffeomorphisms, 4 persistence principle, 38, 100 isometry, 153 Piunikhin, 145 isoperimetric inequality, 30 Poisson bracket, 11 K-area, 86 projective space, 81 K¨ahler,95 pseudo-holomorphic curve, 96 KAM theory, 70 quasi-K¨ahler,95 Kolmogorov, 70 rational Lagrangian submanifold, Lagrangian submanifold, 25 28 Lalonde, 18, 20, 65, 80, 108, 115– reparametrization, 7 117, 126–128, 153, 154 Laudenbach, 76 Salamon, 139, 145 length spectrum, 65 Schwarz, 18, 65, 128, 131, 132, asymptotic, 105 144, 145 Lerman, 85 second variation formula, 117 Lie algebra of Ham(M, Ω), 11 Seidel, 108 Liouville Siburg, 73, 76, 128 class, 27 Sikorav, 3, 28, 39, 72, 76, 128 manifold, 65 , 13 166 INDEX skew product, 103 Slimowitz, 128 stable Lagrangian intersections, 57 Sternberg, 85 strictly ergodic, 104 support, 4 suspension, 26 symplectic connection, 84 curvature of, 84 symplectic fibration, 80 symplectic manifold, 5 symplectomorphism, 9

Ustilovsky, 111, 117, 118 variation, 111 Viterbo, 18, 23, 30, 145 volume, 5

Weinstein, 82

Zehnder, 152 Index

Abreu, 62, 105 Darboux theorem, 5, 151 adjoint action, 11 displacement energy, 19 Akveld, v of Lagrangian submanifolds, almost complex structure, 94 28 Arnold, 2, 70, 152 asymptotic growth, 72 Eliashberg, 18, 20 exact Lagrangian isotopy, 53 Banyaga, 8, 13, 14, 148, 155 exceptional sphere, 94 Bialy, 71, 72, 126, 128 biinvariant (pseudo)distance, 17 fixed extrema, 114 bounded isometry, 153 flatness phenomenon, 128 Floer, 57, 58, 63, 133 Calabi, 148 Floer homology, 131 canonical coordinates, 5 flow, 5 Chekanov, 18, 30 flux closed orbits, 127 conjecture, 10 compactness theorem, 38, 99 group, 148 compatible, 95 of a loop, 148 Conley, 152 Fukaya, 145 coupling class, 85 geodesic, 113 construction, 85 minimal, 80 form, 85 locally, 126 curve shortening, 76, 106 non-degenerate, 116 cusp stable, 116 curve, 100 strictly minimal, 72 solution, 38 locally, 125 cut off, 20 Gromov, 18, 28, 38, 57, 62, 76, cylindrical capacity, 29 86, 95, 99

167 168 INDEX

Guillemin, 85 theorem, 2 Hamiltonian Marsden, 82 diffeomorphism, 7 Mather, 76 equation, 6 Mazur-Ulam theorem, 155 flow, 7 McDuff, 18, 20, 62, 65, 80, 105, function, 6 108, 115–117, 126–128, loop, 46 153 path, 8 mechanical motion, 2 system, 1 Morse homology, 133 vector field, 5 Moser, 63, 70, 74, 82, 97 Herman, 70, 152 Hofer, i, 18–20, 29, 72, 128, 139, non-squeezing, 29 145 normalized Hamiltonian, 6 Hofer’s metric, 23 Novikov, 145 non degeneracy of, 29 Oh, 18, 30, 58, 145 Hofer’s norm, 23 homologically essential, 135 path of diffeomorphisms, 4 persistence principle, 38, 100 isometry, 153 Piunikhin, 145 isoperimetric inequality, 30 Poisson bracket, 11 K-area, 86 projective space, 81 K¨ahler,95 pseudo-holomorphic curve, 96 KAM theory, 70 quasi-K¨ahler,95 Kolmogorov, 70 rational Lagrangian submanifold, Lagrangian submanifold, 25 28 Lalonde, 18, 20, 65, 80, 108, 115– reparametrization, 7 117, 126–128, 153, 154 Laudenbach, 76 Salamon, 139, 145 length spectrum, 65 Schwarz, 18, 65, 128, 131, 132, asymptotic, 105 144, 145 Lerman, 85 second variation formula, 117 Lie algebra of Ham(M, Ω), 11 Seidel, 108 Liouville Siburg, 73, 76, 128 class, 27 Sikorav, 3, 28, 39, 72, 76, 128 manifold, 65 simple group, 13 INDEX 169 skew product, 103 Slimowitz, 128 stable Lagrangian intersections, 57 Sternberg, 85 strictly ergodic, 104 support, 4 suspension, 26 symplectic connection, 84 curvature of, 84 symplectic fibration, 80 symplectic manifold, 5 symplectomorphism, 9

Ustilovsky, 111, 117, 118 variation, 111 Viterbo, 18, 23, 30, 145 volume, 5

Weinstein, 82

Zehnder, 152